feat: huoyi M20 smooth/bridge training and stability docs

Current code snapshot: smooth/bridge task configs, smooth random terrain,
roll/pitch stability rewards, M20 URDF updates, eval profile and stability
training documentation.

scripts/reinforcement_learning/rl_utils.py

@@ -0,0 +1,44 @@
+# Copyright (c) 2025 Deep Robotics
+# SPDX-License-Identifier: BSD 3-Clause
+
+# Copyright (c) 2024-2025 Ziqi Fan
+# SPDX-License-Identifier: Apache-2.0
+
+import torch
+
+import isaaclab.utils.math as math_utils
+
+
+def camera_follow(env):
+    if not hasattr(camera_follow, "smooth_camera_positions"):
+        camera_follow.smooth_camera_positions = []
+    robot_pos = env.unwrapped.scene["robot"].data.root_pos_w[0]
+    robot_quat = env.unwrapped.scene["robot"].data.root_quat_w[0]
+    camera_offset = torch.tensor([-3.0, 0.0, 0.5], dtype=torch.float32, device=env.device)
+    camera_pos = math_utils.transform_points(
+        camera_offset.unsqueeze(0), pos=robot_pos.unsqueeze(0), quat=robot_quat.unsqueeze(0)
+    ).squeeze(0)
+    # camera_pos[2] = torch.clamp(camera_pos[2], min=0.1)
+    window_size = 50
+    camera_follow.smooth_camera_positions.append(camera_pos)
+    if len(camera_follow.smooth_camera_positions) > window_size:
+        camera_follow.smooth_camera_positions.pop(0)
+    smooth_camera_pos = torch.mean(torch.stack(camera_follow.smooth_camera_positions), dim=0)
+    env.unwrapped.viewport_camera_controller.set_view_env_index(env_index=0)
+    env.unwrapped.viewport_camera_controller.update_view_location(
+        eye=smooth_camera_pos.cpu().numpy(), lookat=robot_pos.cpu().numpy()
+    )
+
+
+
+
+import os
+import pickle
+from typing import Any
+
+def reindex_feet(vec):
+    return vec[:, [1, 0, 3, 2]]
+
+def reindex(vec):
+    return vec[:, [3, 4, 5, 0, 1, 2, 9, 10, 11, 6, 7, 8]]
+

scripts/reinforcement_learning/rsl_rl/cli_args.py

@@ -0,0 +1,97 @@
+# Copyright (c) 2025 Deep Robotics
+# SPDX-License-Identifier: BSD 3-Clause
+
+# Copyright (c) 2024-2025 Ziqi Fan
+# SPDX-License-Identifier: Apache-2.0
+
+# Copyright (c) 2024-2025, The Isaac Lab Project Developers.
+# All rights reserved.
+#
+# SPDX-License-Identifier: Apache-2.0
+
+from __future__ import annotations
+
+import argparse
+import random
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from isaaclab_rl.rsl_rl import RslRlOnPolicyRunnerCfg
+
+
+def add_rsl_rl_args(parser: argparse.ArgumentParser):
+    """Add RSL-RL arguments to the parser.
+
+    Args:
+        parser: The parser to add the arguments to.
+    """
+    # create a new argument group
+    arg_group = parser.add_argument_group("rsl_rl", description="Arguments for RSL-RL agent.")
+    # -- experiment arguments
+    arg_group.add_argument(
+        "--experiment_name", type=str, default=None, help="Name of the experiment folder where logs will be stored."
+    )
+    arg_group.add_argument("--run_name", type=str, default=None, help="Run name suffix to the log directory.")
+    # -- load arguments
+    arg_group.add_argument("--resume", action="store_true", default=False, help="Whether to resume from a checkpoint.")
+    arg_group.add_argument("--load_run", type=str, default=None, help="Name of the run folder to resume from.")
+    arg_group.add_argument("--checkpoint", type=str, default=None, help="Checkpoint file to resume from.")
+    # -- logger arguments
+    arg_group.add_argument(
+        "--logger", type=str, default=None, choices={"wandb", "tensorboard", "neptune"}, help="Logger module to use."
+    )
+    arg_group.add_argument(
+        "--log_project_name", type=str, default=None, help="Name of the logging project when using wandb or neptune."
+    )
+
+
+def parse_rsl_rl_cfg(task_name: str, args_cli: argparse.Namespace) -> RslRlOnPolicyRunnerCfg:
+    """Parse configuration for RSL-RL agent based on inputs.
+
+    Args:
+        task_name: The name of the environment.
+        args_cli: The command line arguments.
+
+    Returns:
+        The parsed configuration for RSL-RL agent based on inputs.
+    """
+    from isaaclab_tasks.utils.parse_cfg import load_cfg_from_registry
+
+    # load the default configuration
+    rslrl_cfg: RslRlOnPolicyRunnerCfg = load_cfg_from_registry(task_name, "rsl_rl_cfg_entry_point")
+    rslrl_cfg = update_rsl_rl_cfg(rslrl_cfg, args_cli)
+    return rslrl_cfg
+
+
+def update_rsl_rl_cfg(agent_cfg: RslRlOnPolicyRunnerCfg, args_cli: argparse.Namespace):
+    """Update configuration for RSL-RL agent based on inputs.
+
+    Args:
+        agent_cfg: The configuration for RSL-RL agent.
+        args_cli: The command line arguments.
+
+    Returns:
+        The updated configuration for RSL-RL agent based on inputs.
+    """
+    # override the default configuration with CLI arguments
+    if hasattr(args_cli, "seed") and args_cli.seed is not None:
+        # randomly sample a seed if seed = -1
+        if args_cli.seed == -1:
+            args_cli.seed = random.randint(0, 10000)
+        agent_cfg.seed = args_cli.seed
+    if args_cli.resume is not None:
+        agent_cfg.resume = args_cli.resume
+    if args_cli.load_run is not None:
+        agent_cfg.load_run = args_cli.load_run
+    if args_cli.checkpoint is not None:
+        agent_cfg.load_checkpoint = args_cli.checkpoint
+    if args_cli.run_name is not None:
+        agent_cfg.run_name = args_cli.run_name
+    if args_cli.logger is not None:
+        agent_cfg.logger = args_cli.logger
+    # set the project name for wandb and neptune
+    if agent_cfg.logger in {"wandb", "neptune"} and args_cli.log_project_name:
+        agent_cfg.wandb_project = args_cli.log_project_name
+        agent_cfg.neptune_project = args_cli.log_project_name
+
+    return agent_cfg

scripts/reinforcement_learning/rsl_rl/eval.py

@@ -0,0 +1,716 @@
+# Copyright (c) 2025 Deep Robotics
+# SPDX-License-Identifier: BSD 3-Clause
+
+"""Evaluate an RSL-RL checkpoint with criterion metrics migrated from legged_gym."""
+
+from __future__ import annotations
+
+import argparse
+import copy
+import json
+import os
+import sys
+from datetime import datetime
+
+import numpy as np
+
+PROJECT_ROOT = os.path.abspath(os.path.join(os.path.dirname(__file__), "..", "..", ".."))
+PROJECT_SOURCE_PATH = os.path.join(PROJECT_ROOT, "source", "rl_training")
+if PROJECT_SOURCE_PATH not in sys.path:
+    sys.path.insert(0, PROJECT_SOURCE_PATH)
+
+from isaaclab.app import AppLauncher
+
+sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__), "..")))
+import cli_args
+
+parser = argparse.ArgumentParser(description="Evaluate an RSL-RL checkpoint with criterion metrics.")
+parser.add_argument("--num_envs", type=int, default=250, help="Number of parallel environments.")
+parser.add_argument("--task", type=str, default="Rough-Deeprobotics-M20-v0", help="Name of the task.")
+parser.add_argument(
+    "--agent", type=str, default="rsl_rl_cfg_entry_point", help="Name of the RL agent configuration entry point."
+)
+parser.add_argument("--seed", type=int, default=None, help="Seed used for the environment.")
+parser.add_argument(
+    "--eval_type",
+    type=str,
+    default="all",
+    choices=[
+        "all",
+        "upper_limits",
+        "velocity_error",
+        "access_rate",
+        "time_consumption",
+        "energy_consumption",
+        "motion_stability",
+        "stability_margin",
+        "robustness",
+    ],
+)
+parser.add_argument("--episode_length_s", type=float, default=10.0)
+parser.add_argument("--command_x", type=float, default=1.0)
+parser.add_argument("--command_y", type=float, default=0.0)
+parser.add_argument("--command_yaw", type=float, default=0.0)
+parser.add_argument("--terrain_size", type=float, default=2.0, help="Evaluation terrain tile size in meters.")
+parser.add_argument("--success_distance", type=float, default=2.0)
+parser.add_argument("--start_x", type=float, default=-0.9, help="Evaluation reset x position on the terrain tile.")
+parser.add_argument("--start_y", type=float, default=0.0, help="Evaluation reset y position on the terrain tile.")
+parser.add_argument("--num_levels", type=int, default=8, help="Number of difficulty levels for upper_limits.")
+parser.add_argument("--envs_per_cell", type=int, default=8, help="Number of robots per terrain/difficulty cell.")
+parser.add_argument(
+    "--terrain_type",
+    type=str,
+    default="all",
+    choices=["all", "stairs_up", "stairs_down", "smooth_slope", "rough_slope", "boxes"],
+    help="Terrain type used by upper_limits.",
+)
+parser.add_argument("--tilt_threshold", type=float, default=0.70, help="Tilt threshold in radians for success/stability.")
+parser.add_argument("--acc_threshold", type=float, default=2.0, help="Base acceleration threshold for instability.")
+parser.add_argument("--robot_mass", type=float, default=25.482, help="Robot mass used for COT.")
+parser.add_argument("--output", type=str, default=None, help="Optional output JSON path.")
+parser.add_argument("--print_json", action="store_true", default=False, help="Print the saved JSON payload to stdout.")
+parser.add_argument("--verbose", action="store_true", default=False, help="Print per-terrain per-level debug lines.")
+parser.add_argument(
+    "--include_summary",
+    action="store_true",
+    default=False,
+    help="When eval_type=all, also run a second mixed-terrain rollout for summary metrics.",
+)
+cli_args.add_rsl_rl_args(parser)
+AppLauncher.add_app_launcher_args(parser)
+args_cli, hydra_args = parser.parse_known_args()
+
+sys.argv = [sys.argv[0]] + hydra_args
+
+app_launcher = AppLauncher(args_cli)
+simulation_app = app_launcher.app
+
+import gymnasium as gym
+import torch
+from rsl_rl.runners import OnPolicyRunner
+
+from isaaclab.envs import (
+    DirectMARLEnv,
+    DirectMARLEnvCfg,
+    DirectRLEnvCfg,
+    ManagerBasedRLEnvCfg,
+    multi_agent_to_single_agent,
+)
+from isaaclab.terrains.config.rough import ROUGH_TERRAINS_CFG
+from isaaclab_rl.rsl_rl import RslRlOnPolicyRunnerCfg, RslRlVecEnvWrapper
+from isaaclab_tasks.utils import get_checkpoint_path
+from isaaclab_tasks.utils.hydra import hydra_task_config
+
+import rl_training
+import rl_training.tasks  # noqa: F401
+
+EVAL_TYPES = [
+    "upper_limits",
+    "velocity_error",
+    "access_rate",
+    "time_consumption",
+    "energy_consumption",
+    "motion_stability",
+    "stability_margin",
+    "robustness",
+]
+
+TERRAIN_LEVELS = {
+    "stairs_up": [0.04, 0.06, 0.08, 0.10, 0.12, 0.14, 0.16, 0.18],
+    "stairs_down": [0.04, 0.06, 0.08, 0.10, 0.12, 0.14, 0.16, 0.18],
+    "smooth_slope": [0.05, 0.10, 0.15, 0.20, 0.25, 0.30, 0.35, 0.40],
+    "rough_slope": [0.05, 0.10, 0.15, 0.20, 0.25, 0.30, 0.35, 0.40],
+    "boxes": [0.04, 0.06, 0.08, 0.10, 0.12, 0.14, 0.16, 0.18],
+}
+
+TERRAIN_TO_SUBTERRAIN = {
+    "stairs_up": "pyramid_stairs",
+    "stairs_down": "pyramid_stairs_inv",
+    "smooth_slope": "hf_pyramid_slope",
+    "rough_slope": "random_rough",
+    "boxes": "boxes",
+}
+
+
+def apply_eval_terrain_baseline(env_cfg) -> None:
+    """Use the same evaluation terrain generator for every task."""
+    if env_cfg.scene.terrain.terrain_type == "plane":
+        env_cfg.scene.terrain.terrain_type = "generator"
+    env_cfg.scene.terrain.terrain_generator = copy.deepcopy(ROUGH_TERRAINS_CFG)
+    env_cfg.scene.terrain.max_init_terrain_level = None
+    env_cfg.scene.terrain.terrain_generator.size = (args_cli.terrain_size, args_cli.terrain_size)
+    env_cfg.scene.terrain.terrain_generator.border_width = 2.0
+    env_cfg.scene.terrain.terrain_generator.curriculum = False
+    env_cfg.scene.terrain.terrain_generator.num_rows = 5
+    env_cfg.scene.terrain.terrain_generator.num_cols = 5
+    env_cfg.scene.terrain.terrain_generator.use_cache = False
+    for sub_cfg in env_cfg.scene.terrain.terrain_generator.sub_terrains.values():
+        if hasattr(sub_cfg, "platform_width"):
+            sub_cfg.platform_width = min(0.4, args_cli.terrain_size * 0.2)
+        if hasattr(sub_cfg, "border_width"):
+            sub_cfg.border_width = min(0.1, args_cli.terrain_size * 0.05)
+        if hasattr(sub_cfg, "step_width"):
+            sub_cfg.step_width = min(0.25, args_cli.terrain_size * 0.125)
+        if hasattr(sub_cfg, "grid_width"):
+            sub_cfg.grid_width = min(0.3, args_cli.terrain_size * 0.15)
+
+
+def resolve_checkpoint(log_root_path: str, load_run: str | None, checkpoint: str | None) -> str:
+    if checkpoint is not None:
+        checkpoint_path = os.path.expanduser(checkpoint)
+        if os.path.isabs(checkpoint_path) or os.path.dirname(checkpoint_path):
+            checkpoint_path = os.path.abspath(checkpoint_path)
+            if not os.path.isfile(checkpoint_path):
+                raise FileNotFoundError(f"Unable to find checkpoint: {checkpoint_path}")
+            return checkpoint_path
+
+    if load_run is not None:
+        run_path = os.path.expanduser(load_run)
+        if os.path.isabs(run_path) or os.path.dirname(run_path):
+            run_path = os.path.abspath(run_path)
+            if not os.path.isdir(run_path):
+                raise FileNotFoundError(f"Unable to find run directory: {run_path}")
+            return get_checkpoint_path(os.path.dirname(run_path), os.path.basename(run_path), checkpoint)
+
+    return get_checkpoint_path(log_root_path, load_run, checkpoint)
+
+
+def set_fixed_command(env_cfg, command_x: float, command_y: float, command_yaw: float, episode_length_s: float) -> None:
+    command_cfg = env_cfg.commands.base_velocity
+    command_cfg.heading_command = False
+    command_cfg.resampling_time_range = (episode_length_s + 0.01, episode_length_s + 0.01)
+    command_cfg.ranges.lin_vel_x = (command_x, command_x)
+    command_cfg.ranges.lin_vel_y = (command_y, command_y)
+    command_cfg.ranges.ang_vel_z = (command_yaw, command_yaw)
+
+
+def configure_eval_env(env_cfg, eval_type: str) -> None:
+    env_cfg.scene.num_envs = args_cli.num_envs
+    env_cfg.episode_length_s = args_cli.episode_length_s
+    env_cfg.seed = args_cli.seed if args_cli.seed is not None else 0
+    env_cfg.sim.device = args_cli.device if args_cli.device is not None else env_cfg.sim.device
+    env_cfg.observations.policy.enable_corruption = False
+    set_fixed_command(env_cfg, args_cli.command_x, args_cli.command_y, args_cli.command_yaw, args_cli.episode_length_s)
+
+    apply_eval_terrain_baseline(env_cfg)
+
+    env_cfg.curriculum.terrain_levels = None
+    env_cfg.curriculum.command_levels = None
+
+    for name in (
+        "randomize_rigid_body_material",
+        "randomize_wheel_material",
+        "randomize_support_material",
+        "randomize_rigid_body_mass",
+        "randomize_rigid_body_mass_base",
+        "randomize_rigid_body_inertia",
+        "randomize_com_positions",
+        "randomize_apply_external_force_torque",
+        "randomize_actuator_gains",
+    ):
+        if hasattr(env_cfg.events, name):
+            setattr(env_cfg.events, name, None)
+
+    if hasattr(env_cfg.events, "randomize_reset_base") and env_cfg.events.randomize_reset_base is not None:
+        env_cfg.events.randomize_reset_base.params = {
+            "pose_range": {
+                "x": (args_cli.start_x, args_cli.start_x),
+                "y": (args_cli.start_y, args_cli.start_y),
+                "z": (0.0, 0.0),
+                "roll": (0.0, 0.0),
+                "pitch": (0.0, 0.0),
+                "yaw": (0.0, 0.0),
+            },
+            "velocity_range": {
+                "x": (0.0, 0.0),
+                "y": (0.0, 0.0),
+                "z": (0.0, 0.0),
+                "roll": (0.0, 0.0),
+                "pitch": (0.0, 0.0),
+                "yaw": (0.0, 0.0),
+            },
+        }
+
+    if hasattr(env_cfg.events, "randomize_push_robot"):
+        env_cfg.events.randomize_push_robot = None
+    if eval_type == "robustness" and hasattr(env_cfg.events, "randomize_push_robot"):
+        # Keep robustness deterministic enough to compare runs, but enable interval pushes.
+        from isaaclab.managers import EventTermCfg as EventTerm
+        import rl_training.tasks.manager_based.locomotion.velocity.mdp as mdp
+
+        env_cfg.events.randomize_push_robot = EventTerm(
+            func=mdp.push_by_setting_velocity,
+            mode="interval",
+            interval_range_s=(2.0, 2.01),
+            params={"velocity_range": {"x": (-0.5, 0.5), "y": (-0.5, 0.5)}},
+        )
+
+
+def configure_upper_limit_env(env_cfg, terrain_types: list[str]) -> None:
+    """Configure one fixed terrain grid: rows are difficulty levels, columns are terrain types."""
+    configure_eval_env(env_cfg, "all")
+    terrain_generator = env_cfg.scene.terrain.terrain_generator
+    terrain_generator.curriculum = True
+    terrain_generator.num_rows = args_cli.num_levels
+    terrain_generator.num_cols = len(terrain_types)
+    terrain_generator.difficulty_range = (0.0, 1.0)
+    if hasattr(terrain_generator, "seed"):
+        terrain_generator.seed = args_cli.seed if args_cli.seed is not None else 0
+    env_cfg.scene.terrain.max_init_terrain_level = args_cli.num_levels - 1
+    env_cfg.scene.num_envs = args_cli.num_levels * len(terrain_types) * args_cli.envs_per_cell
+
+    selected_subterrains = [TERRAIN_TO_SUBTERRAIN[terrain_type] for terrain_type in terrain_types]
+    original_subterrains = terrain_generator.sub_terrains
+    terrain_generator.sub_terrains = {
+        subterrain_name: copy.deepcopy(original_subterrains[subterrain_name]) for subterrain_name in selected_subterrains
+    }
+    for name, sub_cfg in terrain_generator.sub_terrains.items():
+        sub_cfg.proportion = selected_subterrains.count(name) / len(selected_subterrains)
+
+    for terrain_type in terrain_types:
+        sub_cfg = terrain_generator.sub_terrains[TERRAIN_TO_SUBTERRAIN[terrain_type]]
+        levels = TERRAIN_LEVELS[terrain_type][: args_cli.num_levels]
+        if terrain_type in ("stairs_up", "stairs_down"):
+            sub_cfg.step_height_range = (levels[0], levels[-1])
+        elif terrain_type == "smooth_slope":
+            sub_cfg.slope_range = (levels[0], levels[-1])
+        elif terrain_type == "rough_slope":
+            sub_cfg.noise_range = (levels[0], levels[-1])
+            sub_cfg.noise_step = 0.01
+        elif terrain_type == "boxes":
+            sub_cfg.grid_height_range = (levels[0], levels[-1])
+
+
+def assign_upper_limit_origins(env, terrain_types: list[str]) -> tuple[torch.Tensor, torch.Tensor]:
+    """Assign an equal number of robots to every terrain/difficulty cell."""
+    base_env = env.unwrapped
+    terrain = base_env.scene.terrain
+    device = base_env.device
+    num_levels = args_cli.num_levels
+    num_types = len(terrain_types)
+    levels = []
+    types = []
+    for terrain_type_id in range(num_types):
+        for level_id in range(num_levels):
+            levels.extend([level_id] * args_cli.envs_per_cell)
+            types.extend([terrain_type_id] * args_cli.envs_per_cell)
+    levels_tensor = torch.tensor(levels, dtype=torch.long, device=device)
+    types_tensor = torch.tensor(types, dtype=torch.long, device=device)
+    terrain.terrain_levels = levels_tensor
+    terrain.terrain_types = types_tensor
+    terrain.env_origins = terrain.terrain_origins[levels_tensor, types_tensor]
+    env.reset()
+    return levels_tensor, types_tensor
+
+
+def tilt_from_projected_gravity(projected_gravity_b: torch.Tensor) -> torch.Tensor:
+    z_axis = torch.clamp(-projected_gravity_b[:, 2], min=-1.0, max=1.0)
+    return torch.acos(z_axis)
+
+
+def support_margin_xy(root_xy: torch.Tensor, foot_xy: torch.Tensor) -> torch.Tensor:
+    min_xy = torch.min(foot_xy, dim=1).values
+    max_xy = torch.max(foot_xy, dim=1).values
+    inside_x = (root_xy[:, 0] >= min_xy[:, 0]) & (root_xy[:, 0] <= max_xy[:, 0])
+    inside_y = (root_xy[:, 1] >= min_xy[:, 1]) & (root_xy[:, 1] <= max_xy[:, 1])
+    inside = inside_x & inside_y
+    inside_dist = torch.minimum(
+        torch.minimum(root_xy[:, 0] - min_xy[:, 0], max_xy[:, 0] - root_xy[:, 0]),
+        torch.minimum(root_xy[:, 1] - min_xy[:, 1], max_xy[:, 1] - root_xy[:, 1]),
+    )
+    outside_dx = torch.clamp(torch.maximum(min_xy[:, 0] - root_xy[:, 0], root_xy[:, 0] - max_xy[:, 0]), min=0.0)
+    outside_dy = torch.clamp(torch.maximum(min_xy[:, 1] - root_xy[:, 1], root_xy[:, 1] - max_xy[:, 1]), min=0.0)
+    outside_dist = torch.sqrt(outside_dx**2 + outside_dy**2)
+    return torch.where(inside, inside_dist, -outside_dist)
+
+
+def run_rollout(env, policy, include_per_env: bool = False) -> dict[str, object]:
+    base_env = env.unwrapped
+    robot = base_env.scene["robot"]
+    dt = base_env.step_dt
+    device = base_env.device
+    max_steps = int(base_env.max_episode_length)
+    command = torch.tensor([args_cli.command_x, args_cli.command_y, args_cli.command_yaw], device=device)
+    command_xy = command[:2].unsqueeze(0)
+    command_yaw = command[2]
+    command_dir = torch.tensor([args_cli.command_x, args_cli.command_y], device=device)
+    if torch.linalg.norm(command_dir) < 1e-6:
+        command_dir = torch.tensor([1.0, 0.0], device=device)
+    command_dir = command_dir / torch.linalg.norm(command_dir)
+
+    obs = env.get_observations()
+    start_pos = robot.data.root_pos_w[:, :3].clone()
+    last_lin_vel = robot.data.root_lin_vel_w[:, :3].clone()
+    complete_time = torch.full((base_env.num_envs,), float("nan"), device=device)
+    energy = torch.zeros(base_env.num_envs, device=device)
+    route = torch.zeros(base_env.num_envs, device=device)
+    lin_error_sum = torch.zeros(base_env.num_envs, device=device)
+    yaw_error_sum = torch.zeros(base_env.num_envs, device=device)
+    instability_count = torch.zeros(base_env.num_envs, device=device)
+    max_tilt = torch.zeros(base_env.num_envs, device=device)
+    stable_time = torch.zeros(base_env.num_envs, device=device)
+    margin_sum = torch.zeros(base_env.num_envs, device=device)
+    margin_count = 0
+    max_progress = torch.zeros(base_env.num_envs, device=device)
+    reached = torch.zeros(base_env.num_envs, dtype=torch.bool, device=device)
+    ever_fallen = torch.zeros(base_env.num_envs, dtype=torch.bool, device=device)
+    support_steps = torch.zeros(base_env.num_envs, device=device)
+    support_events = torch.zeros(base_env.num_envs, device=device)
+    max_support_time = torch.zeros(base_env.num_envs, device=device)
+    previous_support_active = torch.zeros(base_env.num_envs, dtype=torch.bool, device=device)
+
+    wheel_body_ids, _ = robot.find_bodies(".*_wheel")
+    contact_sensor = base_env.scene.sensors["contact_forces"]
+    support_body_ids, _ = contact_sensor.find_bodies(["base_link", ".*_knee"])
+
+    for step in range(max_steps):
+        pre_progress = torch.sum((robot.data.root_pos_w[:, :2] - start_pos[:, :2]) * command_dir.unsqueeze(0), dim=1)
+        max_progress = torch.maximum(max_progress, pre_progress)
+        reached |= pre_progress >= args_cli.success_distance
+
+        with torch.inference_mode():
+            actions = policy(obs)
+            if isinstance(actions, tuple):
+                actions = actions[0]
+            obs, _, dones, _ = env.step(actions)
+
+        lin_vel_b = robot.data.root_lin_vel_b[:, :2]
+        yaw_vel_b = robot.data.root_ang_vel_b[:, 2]
+        lin_error_sum += torch.linalg.norm(lin_vel_b - command_xy, dim=1)
+        yaw_error_sum += torch.abs(yaw_vel_b - command_yaw)
+
+        joint_power = torch.abs(robot.data.applied_torque * robot.data.joint_vel)
+        energy += torch.sum(joint_power, dim=1) * dt
+        route += torch.linalg.norm(robot.data.root_lin_vel_w[:, :2], dim=1) * dt
+
+        acc = torch.linalg.norm(robot.data.root_lin_vel_w[:, :3] - last_lin_vel, dim=1) / dt
+        instability_count += (acc > args_cli.acc_threshold).float()
+        last_lin_vel = robot.data.root_lin_vel_w[:, :3].clone()
+
+        tilt = tilt_from_projected_gravity(robot.data.projected_gravity_b)
+        max_tilt = torch.maximum(max_tilt, tilt)
+        stable_time += (tilt < args_cli.tilt_threshold).float() * dt
+
+        progress = torch.sum((robot.data.root_pos_w[:, :2] - start_pos[:, :2]) * command_dir.unsqueeze(0), dim=1)
+        max_progress = torch.maximum(max_progress, progress)
+        reached |= progress >= args_cli.success_distance
+        newly_success = torch.isnan(complete_time) & (progress >= args_cli.success_distance)
+        complete_time[newly_success] = (step + 1) * dt
+        ever_fallen |= tilt > args_cli.tilt_threshold
+
+        support_force = torch.linalg.norm(
+            contact_sensor.data.net_forces_w[:, support_body_ids, :], dim=-1
+        )
+        support_time = contact_sensor.data.current_contact_time[:, support_body_ids]
+        support_active = torch.any(support_force > 15.0, dim=1)
+        support_steps += support_active.float()
+        support_events += (support_active & ~previous_support_active).float()
+        max_support_time = torch.maximum(max_support_time, torch.amax(support_time, dim=1))
+        previous_support_active = support_active
+
+        if len(wheel_body_ids) >= 3:
+            foot_xy = robot.data.body_pos_w[:, wheel_body_ids, :2]
+            margin_sum += support_margin_xy(robot.data.root_pos_w[:, :2], foot_xy)
+            margin_count += 1
+
+    tilt = tilt_from_projected_gravity(robot.data.projected_gravity_b)
+    success = reached & ~ever_fallen
+    success_count = torch.sum(success).item()
+    success_den = max(base_env.num_envs, 1)
+    successful_time = complete_time[success & torch.isfinite(complete_time)]
+    avg_time = torch.mean(successful_time).item() if successful_time.numel() > 0 else float("nan")
+
+    cot = energy / ((route + 1e-6) * args_cli.robot_mass * 9.81)
+    metrics = {
+        "num_envs": float(base_env.num_envs),
+        "episode_length_s": float(args_cli.episode_length_s),
+        "terrain_size_m": float(args_cli.terrain_size),
+        "success_distance_m": float(args_cli.success_distance),
+        "success_count": float(success_count),
+        "success_rate": float(success_count / success_den),
+        "avg_final_progress_m": float(torch.mean(max_progress).item()),
+        "avg_time_consumption_s": float(avg_time),
+        "time_efficiency": float(avg_time / (args_cli.success_distance / max(abs(args_cli.command_x), 1e-6)))
+        if successful_time.numel() > 0
+        else float("nan"),
+        "mean_lin_vel_error": float(torch.mean(lin_error_sum / max_steps).item()),
+        "mean_yaw_vel_error": float(torch.mean(yaw_error_sum / max_steps).item()),
+        "avg_energy_j": float(torch.mean(energy).item()),
+        "avg_cost_of_transport": float(torch.mean(cot).item()),
+        "unstability_rate": float(torch.mean(instability_count / max_steps).item()),
+        "avg_stability_margin_m": float(torch.mean(margin_sum / max(margin_count, 1)).item()),
+        "avg_max_tilt_rad": float(torch.mean(max_tilt).item()),
+        "avg_stable_time_s": float(torch.mean(stable_time).item()),
+        "avg_support_contact_ratio": float(torch.mean(support_steps / max_steps).item()),
+        "avg_support_events": float(torch.mean(support_events).item()),
+        "avg_max_support_contact_time_s": float(torch.mean(max_support_time).item()),
+        "successful_runs_with_support_ratio": float(
+            torch.mean((support_events[success] > 0.0).float()).item()
+        )
+        if success_count > 0
+        else float("nan"),
+    }
+    if include_per_env:
+        metrics["_per_env"] = {
+            "success": success.detach().cpu().tolist(),
+            "max_progress": max_progress.detach().cpu().tolist(),
+            "max_tilt": max_tilt.detach().cpu().tolist(),
+        }
+    return metrics
+
+
+def make_env_and_policy(env_cfg, agent_cfg, resume_path: str):
+    env = gym.make(args_cli.task, cfg=env_cfg)
+    if isinstance(env.unwrapped, DirectMARLEnv):
+        env = multi_agent_to_single_agent(env)
+    env = RslRlVecEnvWrapper(env, clip_actions=agent_cfg.clip_actions)
+
+    runner = OnPolicyRunner(env, agent_cfg.to_dict(), log_dir=None, device=agent_cfg.device)
+    runner.load(resume_path)
+    policy = runner.get_inference_policy(device=env.unwrapped.device)
+    return env, policy
+
+
+def print_upper_limit_matrix(title: str, terrain_types: list[str], matrix: list[list[float | None]]) -> None:
+    """Print an upper-limit success matrix with levels as rows and terrain types as columns."""
+    col_width = 14
+    print(f"\n{title}")
+    print("level".ljust(8) + "".join(terrain_type[: col_width - 1].rjust(col_width) for terrain_type in terrain_types))
+    for level_index, row in enumerate(matrix, start=1):
+        cells = []
+        for value in row:
+            cells.append("   -".rjust(col_width) if value is None else f"{value:.2f}".rjust(col_width))
+        print(f"{level_index:<8}" + "".join(cells))
+
+
+def matrix_to_numpy(matrix: list[list[float | None]]) -> np.ndarray:
+    return np.array([[np.nan if value is None else value for value in row] for row in matrix], dtype=np.float32)
+
+
+def print_numpy_matrix(title: str, matrix: list[list[float | None]]) -> None:
+    print(f"{title}:")
+    print(np.array2string(matrix_to_numpy(matrix).round(2), precision=2, suppress_small=True))
+
+
+def print_upper_limit_summary(result: dict[str, object], terrain_types: list[str]) -> None:
+    matrix = result["_matrix"]
+    print_numpy_matrix("success rate", matrix["success_rate"])
+    print_numpy_matrix("success rate accumulated", matrix["success_rate_accumulated"])
+    score = np.array([result[terrain_type]["score"] for terrain_type in terrain_types], dtype=np.float32)
+    print(f"score: {np.array2string(score.round(2), precision=2, suppress_small=True)} of {len(terrain_types)}")
+
+
+def print_rollout_summary(metrics: dict[str, object]) -> None:
+    success_rate = float(metrics["success_rate"])
+    success_count = int(round(float(metrics["success_count"])))
+    avg_time = float(metrics["avg_time_consumption_s"])
+    avg_time_text = "nan" if np.isnan(avg_time) else f"{avg_time:.2f}"
+    print(f"success rate: {success_rate:.2f}")
+    print(f"success count: {success_count} time consumption: {avg_time_text}")
+    print(f"average power consumption: {float(metrics['avg_energy_j']):.2f}")
+    print(f"average cost of transport: {float(metrics['avg_cost_of_transport']):.2f}")
+    print(f"unstability rate: {float(metrics['unstability_rate']) * 100:.1f}%")
+    print(f"average stability margin: {float(metrics['avg_stability_margin_m']):.2f}")
+    print(
+        f"average max tilt: {float(metrics['avg_max_tilt_rad']):.2f} "
+        f"average max time: {float(metrics['avg_stable_time_s']):.2f}"
+    )
+    support_success_ratio = float(metrics["successful_runs_with_support_ratio"])
+    support_success_text = "nan" if np.isnan(support_success_ratio) else f"{support_success_ratio * 100:.1f}%"
+    print(
+        f"support contact ratio: {float(metrics['avg_support_contact_ratio']) * 100:.1f}% "
+        f"events/episode: {float(metrics['avg_support_events']):.2f} "
+        f"max duration: {float(metrics['avg_max_support_contact_time_s']):.2f}s "
+        f"successful runs using support: {support_success_text}"
+    )
+
+
+def print_metric_result(eval_type: str, metrics: dict[str, object]) -> None:
+    if eval_type == "velocity_error":
+        print(f"mean lin vel error: {float(metrics['mean_lin_vel_error']):.2f}")
+        print(f"mean yaw vel error: {float(metrics['mean_yaw_vel_error']):.2f}")
+    elif eval_type == "access_rate":
+        print(f"success rate: {float(metrics['success_rate']):.2f}")
+    elif eval_type == "time_consumption":
+        success_count = int(round(float(metrics["success_count"])))
+        avg_time = float(metrics["avg_time_consumption_s"])
+        avg_time_text = "nan" if np.isnan(avg_time) else f"{avg_time:.2f}"
+        print(f"success count: {success_count} time consumption: {avg_time_text}")
+    elif eval_type == "energy_consumption":
+        print(f"average power consumption: {float(metrics['avg_energy_j']):.2f}")
+        print(f"average cost of transport: {float(metrics['avg_cost_of_transport']):.2f}")
+    elif eval_type == "motion_stability":
+        print(f"unstability rate: {float(metrics['unstability_rate']) * 100:.1f}%")
+    elif eval_type == "stability_margin":
+        print(f"average stability margin: {float(metrics['avg_stability_margin_m']):.2f}")
+    elif eval_type == "robustness":
+        print(
+            f"average max tilt: {float(metrics['avg_max_tilt_rad']):.2f} "
+            f"average max time: {float(metrics['avg_stable_time_s']):.2f}"
+        )
+
+
+def run_upper_limits(env_cfg, agent_cfg, resume_path: str) -> dict[str, object]:
+    terrain_types = list(TERRAIN_LEVELS) if args_cli.terrain_type == "all" else [args_cli.terrain_type]
+    eval_cfg = copy.deepcopy(env_cfg)
+    configure_upper_limit_env(eval_cfg, terrain_types)
+    env, policy = make_env_and_policy(eval_cfg, agent_cfg, resume_path)
+    level_ids, type_ids = assign_upper_limit_origins(env, terrain_types)
+    metrics = run_rollout(env, policy, include_per_env=True)
+    env.close()
+
+    per_env = metrics.pop("_per_env")
+    success = torch.tensor(per_env["success"], dtype=torch.float32)
+    max_progress = torch.tensor(per_env["max_progress"], dtype=torch.float32)
+    max_tilt = torch.tensor(per_env["max_tilt"], dtype=torch.float32)
+    level_ids_cpu = level_ids.cpu()
+    type_ids_cpu = type_ids.cpu()
+
+    result: dict[str, object] = {}
+    for terrain_type_id, terrain_type in enumerate(terrain_types):
+        level_values = TERRAIN_LEVELS[terrain_type][: args_cli.num_levels]
+        rates = []
+        level_metrics = []
+        for level_index, level_value in enumerate(level_values):
+            mask = (type_ids_cpu == terrain_type_id) & (level_ids_cpu == level_index)
+            success_count = float(torch.sum(success[mask]).item())
+            total_count = int(torch.sum(mask).item())
+            success_rate = success_count / max(total_count, 1)
+            avg_progress = float(torch.mean(max_progress[mask]).item()) if total_count > 0 else float("nan")
+            avg_tilt = float(torch.mean(max_tilt[mask]).item()) if total_count > 0 else float("nan")
+            rates.append(success_rate)
+            level_metrics.append(
+                {
+                    "level": level_index + 1,
+                    "difficulty": level_value,
+                    "success_rate": success_rate,
+                    "success_count": success_count,
+                    "num_envs": total_count,
+                    "avg_progress_m": avg_progress,
+                    "avg_max_tilt_rad": avg_tilt,
+                }
+            )
+            if args_cli.verbose:
+                print(
+                    f"[EVAL] {terrain_type} level={level_index + 1} difficulty={level_value:.3f} "
+                    f"success_rate={success_rate:.3f}"
+                )
+        accumulated = []
+        running = 1.0
+        for rate in rates:
+            running *= rate
+            accumulated.append(running)
+        score = 0.0
+        for idx, acc_rate in enumerate(accumulated):
+            next_fail = 1.0 - rates[idx + 1] if idx + 1 < len(rates) else 1.0
+            score += acc_rate * (idx + 1) * next_fail
+        result[terrain_type] = {
+            "levels": level_metrics,
+            "success_rate": rates,
+            "success_rate_accumulated": accumulated,
+            "score": score,
+            "max_level": len(level_values),
+            "benchmark": "fixed_rough_terrain_v1",
+            "terrain_size_m": args_cli.terrain_size,
+            "success_distance_m": args_cli.success_distance,
+            "episode_length_s": args_cli.episode_length_s,
+        }
+    max_levels = max((len(result[terrain_type]["success_rate"]) for terrain_type in terrain_types), default=0)
+    success_matrix: list[list[float | None]] = []
+    accumulated_matrix: list[list[float | None]] = []
+    difficulty_matrix: list[list[float | None]] = []
+    for level_index in range(max_levels):
+        success_row = []
+        accumulated_row = []
+        difficulty_row = []
+        for terrain_type in terrain_types:
+            rates = result[terrain_type]["success_rate"]
+            accumulated = result[terrain_type]["success_rate_accumulated"]
+            levels = result[terrain_type]["levels"]
+            success_row.append(rates[level_index] if level_index < len(rates) else None)
+            accumulated_row.append(accumulated[level_index] if level_index < len(accumulated) else None)
+            difficulty_row.append(levels[level_index]["difficulty"] if level_index < len(levels) else None)
+        success_matrix.append(success_row)
+        accumulated_matrix.append(accumulated_row)
+        difficulty_matrix.append(difficulty_row)
+
+    result["_matrix"] = {
+        "terrain_types": terrain_types,
+        "success_rate": success_matrix,
+        "success_rate_accumulated": accumulated_matrix,
+        "difficulty": difficulty_matrix,
+        "terrain_size_m": args_cli.terrain_size,
+        "success_distance_m": args_cli.success_distance,
+        "episode_length_s": args_cli.episode_length_s,
+    }
+    result["_summary"] = metrics
+    print_upper_limit_summary(result, terrain_types)
+    return result
+
+
+def filter_metrics(metrics: dict[str, float], eval_type: str) -> dict[str, float]:
+    if eval_type == "all":
+        return metrics
+    keys = {
+        "velocity_error": ("mean_lin_vel_error", "mean_yaw_vel_error"),
+        "access_rate": ("success_count", "success_rate", "avg_final_progress_m"),
+        "time_consumption": ("success_count", "avg_time_consumption_s", "time_efficiency"),
+        "energy_consumption": ("avg_energy_j", "avg_cost_of_transport"),
+        "motion_stability": ("unstability_rate",),
+        "stability_margin": ("avg_stability_margin_m",),
+        "robustness": ("success_rate", "avg_max_tilt_rad", "avg_stable_time_s"),
+    }[eval_type]
+    return {key: metrics[key] for key in keys}
+
+
+@hydra_task_config(args_cli.task, args_cli.agent)
+def main(env_cfg: ManagerBasedRLEnvCfg | DirectRLEnvCfg | DirectMARLEnvCfg, agent_cfg: RslRlOnPolicyRunnerCfg):
+    agent_cfg = cli_args.update_rsl_rl_cfg(agent_cfg, args_cli)
+
+    log_root_path = os.path.abspath(os.path.join("logs", "rsl_rl", agent_cfg.experiment_name))
+    resume_path = resolve_checkpoint(log_root_path, agent_cfg.load_run, agent_cfg.load_checkpoint)
+    print(f"[INFO] Evaluating checkpoint: {resume_path}")
+
+    if args_cli.eval_type == "upper_limits":
+        selected = run_upper_limits(env_cfg, agent_cfg, resume_path)
+    elif args_cli.eval_type == "all":
+        upper_limits = run_upper_limits(env_cfg, agent_cfg, resume_path)
+        selected = {"upper_limits": upper_limits}
+        print_rollout_summary(upper_limits["_summary"])
+        if args_cli.include_summary:
+            summary_cfg = copy.deepcopy(env_cfg)
+            configure_eval_env(summary_cfg, "all")
+            env, policy = make_env_and_policy(summary_cfg, agent_cfg, resume_path)
+            summary = run_rollout(env, policy)
+            env.close()
+            selected["summary"] = summary
+    else:
+        eval_type_for_env = "robustness" if args_cli.eval_type == "robustness" else "all"
+        configure_eval_env(env_cfg, eval_type_for_env)
+        env, policy = make_env_and_policy(env_cfg, agent_cfg, resume_path)
+        metrics = run_rollout(env, policy)
+        selected = filter_metrics(metrics, args_cli.eval_type)
+        env.close()
+        print_metric_result(args_cli.eval_type, selected)
+    if args_cli.print_json:
+        print(json.dumps(selected, indent=2, sort_keys=True))
+
+    output = args_cli.output
+    if output is None:
+        eval_dir = os.path.join(os.path.dirname(resume_path), "eval")
+        os.makedirs(eval_dir, exist_ok=True)
+        suffix = datetime.now().strftime("%Y-%m-%d_%H-%M-%S")
+        output = os.path.join(eval_dir, f"{args_cli.eval_type}_{suffix}.json")
+    else:
+        output = os.path.abspath(os.path.expanduser(output))
+        output_dir = os.path.dirname(output)
+        if output_dir:
+            os.makedirs(output_dir, exist_ok=True)
+    with open(output, "w", encoding="utf-8") as f:
+        json.dump({"eval_type": args_cli.eval_type, "checkpoint": resume_path, "metrics": selected}, f, indent=2)
+    print(f"[INFO] Saved evaluation result to: {output}")
+
+
+if __name__ == "__main__":
+    main()
+    simulation_app.close()

scripts/reinforcement_learning/rsl_rl/eval.sh

@@ -0,0 +1,23 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+TASK=${TASK:-Rough-Deeprobotics-M20-v0}
+CHECKPOINT=${1:-${CHECKPOINT:-}}
+NUM_ENVS=${NUM_ENVS:-250}
+DEVICE=${DEVICE:-cuda:0}
+EVAL_TYPE=${EVAL_TYPE:-all}
+
+if [[ -z "${CHECKPOINT}" ]]; then
+  echo "Usage: $0 <checkpoint.pt>"
+  echo "Or set CHECKPOINT=/path/to/model.pt"
+  exit 1
+fi
+
+echo "Running eval_type=${EVAL_TYPE}"
+python scripts/reinforcement_learning/rsl_rl/eval.py \
+  --task="${TASK}" \
+  --checkpoint="${CHECKPOINT}" \
+  --num_envs="${NUM_ENVS}" \
+  --device="${DEVICE}" \
+  --eval_type="${EVAL_TYPE}" \
+  --headless

scripts/reinforcement_learning/rsl_rl/eval_suite.py

@@ -0,0 +1,893 @@
+# Copyright (c) 2025 Deep Robotics
+# SPDX-License-Identifier: BSD 3-Clause
+
+"""评价套件的命令行入口。
+
+与 ``eval.py`` 并存、互不影响：本脚本不导入也不修改 ``eval.py``，两者可以并行使用，
+输出 JSON 结构相似（指标名完全一致），便于互相对照。
+
+用法示例::
+
+    # 标准评估（默认协议与火蚁描述）
+    python scripts/reinforcement_learning/rsl_rl/eval_suite.py \\
+        --task Rough-Deeprobotics-M20-v0 \\
+        --checkpoint /绝对路径/model_19999.pt \\
+        --headless
+
+    # 上限评估，5 个随机种子，顺带把轨迹存下来
+    python scripts/reinforcement_learning/rsl_rl/eval_suite.py \\
+        --task Rough-Deeprobotics-M20-v0 \\
+        --checkpoint /绝对路径/model_19999.pt \\
+        --protocol standard_v1_multiseed --save-trace /tmp/traces --headless
+
+    # 抗扰评估
+    python scripts/reinforcement_learning/rsl_rl/eval_suite.py \\
+        --task Rough-Deeprobotics-M20-v0 \\
+        --checkpoint /绝对路径/model_19999.pt \\
+        --eval-type robustness --headless
+
+    # 用 WebRTC 看画面（不需要显示器；局域网内用 Isaac Sim WebRTC Streaming Client 连）
+    python scripts/reinforcement_learning/rsl_rl/eval_suite.py \\
+        --task Rough-Deeprobotics-M20-v0 \\
+        --checkpoint /绝对路径/model_19999.pt \\
+        --eval-type upper_limits --terrain-type stairs_up \\
+        --livestream 2
+
+    ``--livestream 1`` 走公网（NVIDIA NVCF 信令，需要 ``PUBLIC_IP``），
+    ``--livestream 2`` 走局域网直连；两者都由 ``AppLauncher`` 提供，详见
+    ``isaaclab.app.AppLauncher`` 的文档。本脚本只额外保证：视口相机跟着被评估的
+    机器人（``--view-env``）、开始 rollout 前留出连接时间（``--pause-s``）、
+    以及把「客户端该连哪里」打印出来。
+
+    协议预设与机器人描述都在 ``eval_kit/`` 里：改口径改 ``protocols.py``，
+    加机器人改 ``profiles.py``。
+"""
+
+from __future__ import annotations
+
+import argparse
+import copy
+import json
+import os
+import subprocess
+import sys
+import tempfile
+from dataclasses import replace
+from typing import Any
+
+#: 原始命令行，用于在需要多个环境时 fork 子进程（见文件末尾的说明）。
+_ORIGINAL_ARGV = list(sys.argv[1:])
+
+# --- 路径引导：必须在导入 eval_kit / rl_training 之前完成 --------------------------
+_HERE = os.path.dirname(os.path.abspath(__file__))
+PROJECT_ROOT = os.path.abspath(os.path.join(_HERE, "..", "..", ".."))
+PROJECT_SOURCE_PATH = os.path.join(PROJECT_ROOT, "source", "rl_training")
+for path in (PROJECT_SOURCE_PATH, PROJECT_ROOT):
+    if path not in sys.path:
+        sys.path.insert(0, path)
+
+from isaaclab.app import AppLauncher  # noqa: E402
+
+from eval_kit.core.report import (  # noqa: E402
+    EVAL_TYPES,
+    filter_metrics,
+    format_aggregate_summary,
+    format_metric_result,
+    format_rollout_summary,
+    format_upper_limit_summary,
+)
+from eval_kit.core.schema import EvalProtocol  # noqa: E402
+from eval_kit.core.terrain_spec import get_terrain_spec  # noqa: E402
+from eval_kit.profiles import PROFILES, get_profile  # noqa: E402
+from eval_kit.protocols import PROTOCOLS, get_protocol  # noqa: E402
+
+
+def _build_parser() -> argparse.ArgumentParser:
+    parser = argparse.ArgumentParser(
+        description="基于 eval_kit 的机器人性能评估（与 eval.py 并存，指标口径一致）。",
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter,
+    )
+    parser.add_argument("--task", type=str, default="Rough-Deeprobotics-M20-v0", help="任务名。")
+    parser.add_argument(
+        "--agent", type=str, default="rsl_rl_cfg_entry_point", help="agent 配置入口名（配合 Hydra）。"
+    )
+    parser.add_argument(
+        "--profile", type=str, default="solenopsis", choices=sorted(PROFILES), help="机器人形态描述。"
+    )
+    parser.add_argument(
+        "--protocol", type=str, default="standard_v1", choices=sorted(PROTOCOLS), help="评估协议预设。"
+    )
+    parser.add_argument(
+        "--eval-type", dest="eval_type", type=str, default="all", choices=list(EVAL_TYPES), help="评估类型。"
+    )
+
+    checkpoint = parser.add_argument_group("checkpoint")
+    checkpoint.add_argument("--checkpoint", type=str, default=None, help="checkpoint 路径（支持绝对路径）。")
+    checkpoint.add_argument("--load-run", dest="load_run", type=str, default=None, help="运行目录或目录名。")
+    checkpoint.add_argument(
+        "--experiment-name", dest="experiment_name", type=str, default=None,
+        help="实验名，用于在 logs/rsl_rl/<实验名> 下查找；不传则用 agent 配置里的。",
+    )
+
+    run = parser.add_argument_group("运行规模（不传则取协议预设值）")
+    run.add_argument(
+        "--num-envs", dest="num_envs", type=int, default=250,
+        help="并行环境数（上限评估时会按网格规模自动覆盖）。",
+    )
+    # 注意：--device 由 AppLauncher.add_app_launcher_args() 提供，这里不能再加。
+    # 它保留了 device / headless / enable_cameras / experience / livestream /
+    # rendering_mode / xr 这几个名字，重复定义会直接抛 ValueError。
+    run.add_argument(
+        "--seeds", type=int, nargs="+", default=None,
+        help="重复试验的随机种子，覆盖协议预设；给多个种子时会输出 mean±std。",
+    )
+    run.add_argument("--save-trace", dest="save_trace", type=str, default=None, help="把轨迹存到该目录。")
+    run.add_argument(
+        "--view-env", dest="view_env", type=int, default=None,
+        help="让视口相机跟随第 N 个环境的机器人（非 headless 或开启推流时有效）。"
+             "不传时默认跟随 env 0。环境编号 0 起，"
+             "编号与（地形, 难度）的对应关系会在启动时打印。",
+    )
+    run.add_argument(
+        "--free-camera", dest="free_camera", action="store_true", default=False,
+        help="不要让视口相机跟随任何机器人：起始摆在俯视全场的位置，之后在 WebRTC 客户端里"
+             "自己移动（脚本与 IsaacLab 都不会再改写相机）。同时给了 --view-env 时以本参数为准。",
+    )
+    run.add_argument(
+        "--pause-s", dest="pause_s", type=float, default=None,
+        help="开始 rollout 之前停顿的秒数：期间画面照常推流、物理冻结，按回车可提前开始。"
+             "用来给 WebRTC 客户端留连接时间。开启推流时默认 30 s，其余情况默认 0（不停顿）。",
+    )
+    run.add_argument(
+        "--record-frames", dest="record_frames", type=str, default=None,
+        help="把 rollout 期间的视口画面按帧存成 PNG 到该目录（例如 frame_000150.png）。"
+             "推流传输不可用时用它离线看评估过程；抓帧间隔见 --record-every。",
+    )
+    run.add_argument(
+        "--record-every", dest="record_every", type=int, default=5,
+        help="--record-frames 每多少步抓一帧。",
+    )
+
+    overrides = parser.add_argument_group("口径覆盖（不传则取协议预设值，改动会反映到 fingerprint）")
+    overrides.add_argument("--episode-length-s", dest="episode_length_s", type=float, default=None)
+    overrides.add_argument("--command-x", dest="command_x", type=float, default=None)
+    overrides.add_argument("--command-y", dest="command_y", type=float, default=None)
+    overrides.add_argument("--command-yaw", dest="command_yaw", type=float, default=None)
+    overrides.add_argument("--start-x", dest="start_x", type=float, default=None)
+    overrides.add_argument("--start-y", dest="start_y", type=float, default=None)
+    overrides.add_argument(
+        "--start-z-offset", dest="start_z_offset", type=float, default=None,
+        help="起始高度相对地形瓦片中心的抬升 [m]。倒金字塔台阶这类「中心最低」的地形需要正值，"
+             "否则机器人会被生成在地形内部（默认 0.0，与旧 eval.py 一致）。",
+    )
+    overrides.add_argument("--terrain-size", dest="terrain_size", type=float, default=None, help="地形瓦片边长 [m]。")
+    overrides.add_argument("--success-distance", dest="success_distance", type=float, default=None)
+    overrides.add_argument("--tilt-threshold", dest="tilt_threshold", type=float, default=None, help="摔倒倾角 [rad]。")
+    overrides.add_argument(
+        "--acc-threshold-g", dest="acc_threshold_g", type=float, default=None,
+        help="不稳定判定的加速度阈值，单位 g。",
+    )
+    overrides.add_argument(
+        "--energy-convention", dest="energy_convention", type=str, default=None, choices=["absolute", "signed"]
+    )
+    overrides.add_argument(
+        "--robot-mass", dest="robot_mass", type=float, default=None,
+        help="覆盖 COT 用质量；默认从资产自动读取。",
+    )
+    overrides.add_argument(
+        "--terrain-type", dest="terrain_type", type=str, default="all",
+        help="上限评估只用某一种地形：all 或地形规格里的类型名。",
+    )
+    overrides.add_argument("--num-levels", dest="num_levels", type=int, default=None, help="上限评估的难度等级数。")
+    overrides.add_argument("--envs-per-cell", dest="envs_per_cell", type=int, default=None, help="每格机器人数。")
+    overrides.add_argument(
+        "--no-terrain-bounds", dest="enforce_terrain_bounds", action="store_false", default=None,
+        help="去掉越界终止项。地砖小于 3.8 m 时它会误判最外侧格子（详见 --help 里 protocol 的说明）。",
+    )
+
+    output = parser.add_argument_group("输出")
+    output.add_argument("--output", type=str, default=None, help="输出 JSON 路径；默认写在 checkpoint 同目录的 eval_kit/ 下。")
+    output.add_argument("--print-json", dest="print_json", action="store_true", default=False, help="把 JSON 打到 stdout。")
+    # 不能叫 --verbose：AppLauncher.add_app_launcher_args() 已经占用了这个名字，
+    # 重复定义会让 argparse 直接抛 ArgumentError（旧 eval.py 就卡在这里）。
+    output.add_argument(
+        "--verbose-terrain", dest="verbose_terrain", action="store_true", default=False,
+        help="打印逐格地形/难度调试信息。（注意 --verbose 属于 AppLauncher，行为不同）",
+    )
+    output.add_argument(
+        "--include-summary", dest="include_summary", action="store_true", default=False,
+        help="eval-type=all 时额外跑一次混合地形汇总。",
+    )
+    # 内部参数：一个进程只能建一个环境，需要多个环境时由父进程 fork 自己并传这些参数。
+    # 不建议手工使用。
+    internal = parser.add_argument_group("内部（子进程用，勿手工指定）")
+    internal.add_argument("--stage", type=str, default="", choices=["", "upper", "plain"])
+    internal.add_argument("--stage-seed", dest="stage_seed", type=int, default=0)
+    internal.add_argument("--stage-out", dest="stage_out", type=str, default=None)
+    return parser
+
+
+def _apply_overrides(protocol: EvalProtocol, args: argparse.Namespace) -> EvalProtocol:
+    """把命令行显式给出的口径覆盖到协议预设上。没给的不动，保持预设原值。"""
+    updates: dict[str, object] = {}
+    if args.episode_length_s is not None:
+        updates["steps"] = int(round(args.episode_length_s / protocol.dt))
+    if args.command_x is not None:
+        updates["command"] = (args.command_x, protocol.command[1], protocol.command[2])
+    if args.command_y is not None:
+        current = updates.get("command", protocol.command)
+        updates["command"] = (current[0], args.command_y, current[2])
+    if args.command_yaw is not None:
+        current = updates.get("command", protocol.command)
+        updates["command"] = (current[0], current[1], args.command_yaw)
+    if args.start_x is not None:
+        updates["start_pose"] = (args.start_x, protocol.start_pose[1], protocol.start_pose[2])
+    if args.start_y is not None:
+        current = updates.get("start_pose", protocol.start_pose)
+        updates["start_pose"] = (current[0], args.start_y, current[2])
+    if args.start_z_offset is not None:
+        updates["start_z_offset_m"] = args.start_z_offset
+    if args.terrain_size is not None:
+        updates["terrain_size_m"] = args.terrain_size
+    if args.success_distance is not None:
+        updates["success_distance_m"] = args.success_distance
+    if args.tilt_threshold is not None:
+        updates["tilt_threshold_rad"] = args.tilt_threshold
+    if args.acc_threshold_g is not None:
+        updates["acc_threshold_g"] = args.acc_threshold_g
+    if args.energy_convention is not None:
+        updates["energy_convention"] = args.energy_convention
+    if args.num_levels is not None:
+        updates["num_levels"] = args.num_levels
+    if args.envs_per_cell is not None:
+        updates["envs_per_cell"] = args.envs_per_cell
+    if args.seeds is not None:
+        updates["seeds"] = tuple(args.seeds)
+    if args.enforce_terrain_bounds is not None:
+        updates["enforce_terrain_bounds"] = args.enforce_terrain_bounds
+    return replace(protocol, **updates)  # type: ignore[arg-type]
+
+
+def _terrain_types_for(args: argparse.Namespace, spec) -> list[str]:
+    if args.terrain_type == "all":
+        return list(spec.types)
+    if args.terrain_type not in spec.types:
+        raise SystemExit(
+            f"地形类型 {args.terrain_type!r} 不在规格 {spec.name!r} 里；可选：{list(spec.types)}"
+        )
+    return [args.terrain_type]
+
+
+def _resolve_output_path(args: argparse.Namespace, resume_path: str, eval_type: str) -> str:
+    if args.output is not None:
+        path = os.path.abspath(os.path.expanduser(args.output))
+    else:
+        from datetime import datetime
+
+        directory = os.path.join(os.path.dirname(resume_path), "eval_kit")
+        suffix = datetime.now().strftime("%Y-%m-%d_%H-%M-%S")
+        path = os.path.join(directory, f"{eval_type}_{suffix}.json")
+    parent = os.path.dirname(path)
+    if parent:
+        os.makedirs(parent, exist_ok=True)
+    return path
+
+
+def _livestream_value(args_cli: argparse.Namespace) -> int:
+    """解析实际生效的推流模式：0 关、1 公网 WebRTC、2 局域网 WebRTC。
+
+    ``--livestream`` 的默认值是 -1，含义是「听 ``LIVESTREAM`` 环境变量的」，
+    所以 -1 时要回退去读环境变量，而不是当成 0。
+    """
+    try:
+        value = int(getattr(args_cli, "livestream", -1))
+    except (TypeError, ValueError):
+        value = -1
+    if value < 0:
+        try:
+            value = int(os.environ.get("LIVESTREAM", 0) or 0)
+        except ValueError:
+            value = 0
+    return value
+
+
+def _display_enabled(args_cli: argparse.Namespace) -> bool:
+    """这次运行会不会给用户看画面（本地 GUI 或推流）。"""
+    return not getattr(args_cli, "headless", True) or _livestream_value(args_cli) > 0
+
+
+def _viewer_env_index(args_cli: argparse.Namespace) -> int | None:
+    """视口相机跟随哪个环境；不跟随时（自由视角、或完全不需要画面）返回 None。
+
+    必须在构造 ``AppLauncher`` **之前**调用：``AppLauncher`` 会把 ``livestream``
+    从 namespace 里 pop 掉，还会在开启推流时把 ``headless`` 改写成 True（推流在
+    服务端本来就按 headless 启动），等它跑完再读这两个值就分不清「用户要不要看画面」了。
+
+    推流时 ``SimulationContext`` 的 ``has_gui()``（本地 GUI 或推流 GUI）仍为真、
+    ``env_cfg.viewer`` 也仍然生效，所以推流和本地 GUI 两种情况都要设视口相机，
+    否则推流画面停在默认相机上、看不到机器人。
+    """
+    if not _display_enabled(args_cli) or getattr(args_cli, "free_camera", False):
+        return None
+    return 0 if args_cli.view_env is None else int(args_cli.view_env)
+
+
+def _set_free_camera_overview(env: Any) -> None:
+    """自由视角模式：把相机摆到能看见整片地形的位置。
+
+    ``env_cfg.viewer.origin_type`` 是 ``world`` 时，``ViewportCameraController``
+    只在初始化时设一次相机、之后不再改写，所以这里设的视角会一直保留，
+    客户端里可以随便移动鼠标/键盘来观察。
+    """
+    import numpy as np
+
+    origins = env.unwrapped.scene.env_origins
+    center = origins.mean(dim=0)
+    radius = float((origins - center).norm(dim=1).max().item())
+    center_np = center.detach().cpu().numpy()
+    eye = center_np + np.array([0.0, -1.6 * radius, 1.25 * radius], dtype=center_np.dtype)
+    env.unwrapped.sim.set_camera_view(eye=eye, target=center_np)
+    print(
+        f"[eval_kit] 自由视角：相机不跟随机器人，起始俯视全场"
+        f"（地形中心 {np.round(center_np, 1).tolist()}，半径 {radius:.1f} m），"
+        f"之后在客户端里自己移动",
+        flush=True,
+    )
+
+
+def _allow_client_resize(args_cli: argparse.Namespace) -> None:
+    """让 WebRTC 客户端可以改画面尺寸（默认关掉会冻住画面）。
+
+    客户端一旦改窗口大小/全屏，就会发 ``onCustomResizeMessage``；服务端默认拒绝
+    （``... must both be positive and /app/livestream/allowDynamicResize enabled``），
+    同时推流用的 ``shared swapchain buffer`` 进入坏状态 —— 之后画面不再更新，而仿真
+    照常在跑（终端里的 rollout 进度还在涨）。这两个设置通过 ``--kit_args`` 传给 Kit。
+
+    用户自己显式给了同名设置就不覆盖（想关掉就传 ``--kit_args=--/app/livestream/allowDynamicResize=false``）。
+    """
+    wanted = ("--/app/livestream/allowResize=true", "--/app/livestream/allowDynamicResize=true")
+    existing = args_cli.kit_args or ""
+    extra = [flag for flag in wanted if flag.split("=")[0] not in existing]
+    if not extra:
+        return
+    args_cli.kit_args = " ".join(filter(None, [existing, *extra]))
+    print(
+        f"[eval_kit] 推流追加 kit 设置：{' '.join(extra)}（允许客户端调整窗口大小/全屏）",
+        flush=True,
+    )
+
+
+def _grab_viewport_frame(state: dict, directory: str, every: int) -> None:
+    """每 ``every`` 步向视口要一张 PNG。
+
+    抓帧是异步的（等下一次渲染产出），所以这里不阻塞 rollout：上一张还没写完就跳过
+    这一帧。写完的文件名形如 ``frame_000150.png``，可用 ffmpeg 合成视频。
+    """
+    if every <= 0 or state["step"] % every:
+        return
+    pending = state["pending"]
+    if pending is not None and not os.path.exists(pending):
+        return
+    from omni.kit.viewport.utility import capture_viewport_to_file, get_active_viewport
+
+    path = os.path.join(directory, f"frame_{state['step']:06d}.png")
+    state["pending"] = path
+    capture_viewport_to_file(get_active_viewport(), path)
+
+
+def _with_frame_capture(policy: Any, directory: str, every: int) -> Any:
+    """把策略包一层，顺带按间隔抓视口帧（策略每被调用一次就是 rollout 推进一步）。"""
+    os.makedirs(directory, exist_ok=True)
+    state: dict = {"step": 0, "pending": None}
+    print(f"[eval_kit] rollout 期间抓帧到 {directory}（每 {every} 步一张 PNG）", flush=True)
+
+    def wrapped(obs: Any) -> Any:
+        state["step"] += 1
+        _grab_viewport_frame(state, directory, every)
+        return policy(obs)
+
+    return wrapped
+
+
+def _report_recorded_frames(directory: str, every: int, dt: float) -> None:
+    """抓帧结束后提示怎么合成视频（帧是按仿真时间等间隔抓的，fps = 1/(every*dt)）。"""
+    frames = sorted(name for name in os.listdir(directory) if name.endswith(".png"))
+    if not frames:
+        print(f"[eval_kit] 抓帧目录 {directory} 里没有帧（视口没渲染？）", flush=True)
+        return
+    fps = 1.0 / (every * dt) if every > 0 and dt > 0 else 10.0
+    print(f"[eval_kit] 抓到 {len(frames)} 帧：{directory}", flush=True)
+    print(
+        f"[eval_kit]   合成视频（按仿真速度播放）：cd {directory} && "
+        f"ffmpeg -y -framerate {fps:.2f} -pattern_type glob -i '*.png' "
+        f"-pix_fmt yuv420p -c:v libx264 -vf 'pad=ceil(iw/2)*2:ceil(ih/2)*2' /tmp/eval.mp4",
+        flush=True,
+    )
+
+
+def _local_ipv4_addresses() -> list[str]:
+    """本机的非回环 IPv4 地址，用来告诉用户 WebRTC 客户端该连哪里。"""
+    import socket
+
+    addresses: list[str] = []
+    try:
+        for info in socket.getaddrinfo(socket.gethostname(), None, socket.AF_INET):
+            address = info[4][0]
+            if address not in addresses:
+                addresses.append(address)
+    except OSError:
+        pass
+    try:
+        # 不真的发包：只是让内核按路由表挑一个默认出口地址
+        with socket.socket(socket.AF_INET, socket.SOCK_DGRAM) as probe:
+            probe.connect(("8.8.8.8", 80))
+            address = probe.getsockname()[0]
+            if address not in addresses:
+                addresses.append(address)
+    except OSError:
+        pass
+    return [address for address in addresses if not address.startswith("127.")]
+
+
+def _print_livestream_info(livestream: int) -> None:
+    """打印客户端该连到哪里。要在 Kit 起来之后调用，端口从 carb 设置里读实际值。"""
+    import carb.settings
+
+    port = carb.settings.get_settings().get("/app/livestream/port")
+    addresses = _local_ipv4_addresses()
+    if livestream == 1:
+        public_ip = os.environ.get("PUBLIC_IP") or (addresses[0] if addresses else "127.0.0.1")
+        print(f"[eval_kit] WebRTC 推流已启用（--livestream 1，公网，PUBLIC_IP={public_ip}）", flush=True)
+    else:
+        print("[eval_kit] WebRTC 推流已启用（--livestream 2，局域网直连）", flush=True)
+    if addresses:
+        targets = "  ".join(f"{address}:{port}" for address in addresses)
+        print(f"[eval_kit]   客户端连接地址：{targets}", flush=True)
+    else:
+        print(f"[eval_kit]   客户端连接地址：<本机 IP>:{port}", flush=True)
+    print(
+        "[eval_kit]   本子进程结束后推流会断开（每个阶段/种子都是一个新进程），"
+        "客户端重连即可",
+        flush=True,
+    )
+
+
+def _hold_before_rollout(sim: Any, seconds: float, *, start_height_note: str) -> None:
+    """停顿 ``seconds`` 秒再开始 rollout：画面继续推流、物理冻结。
+
+    用 ``sim.render()`` 而不是 ``time.sleep``：独立进程模式下 Kit 的更新循环由我们
+    调用 ``app.update()`` 驱动，光 sleep 推流画面会冻住。``render()`` 内部会先把
+    ``/app/player/playSimulations`` 关掉再 update，因此画面在刷新、时间不前进。
+    另外把 timeline 暂停，避免物理在等待期间自己走掉（那会让评估起点不是刚重置的状态）。
+    """
+    import select
+    import time as _time
+
+    sim.pause()
+    started = _time.monotonic()
+    deadline = started + seconds
+    # 明确说明「悬空不动」是正常的：这一窗物理是冻结的，机器人正停在法规的起始位姿上
+    print(
+        f"[eval_kit] 画面已就绪，先停 {seconds:.0f} s 再开始 rollout —— 期间物理是冻结的，"
+        f"机器人在起始位姿上不动{start_height_note}，看起来「悬空不动」是正常的。",
+        flush=True,
+    )
+    if sys.stdin is not None and sys.stdin.isatty():
+        print("[eval_kit]   在终端按回车可立即开始 rollout", flush=True)
+    next_report = started + 5.0
+    while _time.monotonic() < deadline:
+        sim.render()
+        if sys.stdin is not None and sys.stdin.isatty() and select.select([sys.stdin], [], [], 0)[0]:
+            sys.stdin.readline()
+            break
+        if _time.monotonic() >= next_report:
+            print(f"[eval_kit]   剩余 {max(0.0, deadline - _time.monotonic()):.0f} s（物理仍冻结）", flush=True)
+            next_report += 5.0
+    sim.play()
+    print(f"[eval_kit] 停顿 {_time.monotonic() - started:.1f} s 结束，开始 rollout（画面从这里才动）", flush=True)
+
+
+def _format_early_reset_warning(early: dict, protocol: EvalProtocol, min_terrain_size: float) -> str:
+    """把早退统计变成可执行的建议。"""
+    import json as _json
+
+    lines = [
+        "[警告] 有环境在 episode 中途被重置，路程/进度类指标可能被污染："
+        f"{_json.dumps(early, ensure_ascii=False)}"
+    ]
+    if protocol.enforce_terrain_bounds:
+        lines.append("  最常见的原因是 IsaacLab 的 terrain_out_of_bounds 终止项：它比较机器人的")
+        lines.append("  世界坐标与整张地图的范围，而不是机器人自己那块地砖，于是最外侧的格子")
+        lines.append(f"  从第 0 步起就被判越界。当前地砖 {protocol.terrain_size_m} m，")
+        lines.append(f"  要保留该终止项需要 >= {min_terrain_size:.1f} m。三种处置：")
+        lines.append(f"    1) --terrain-size {max(min_terrain_size, protocol.terrain_size_m):.1f}  保留终止项，消除误判")
+        lines.append("    2) --no-terrain-bounds         去掉该终止项")
+        lines.append("    3) --protocol standard_v1_clean  等价于方案 1")
+    return "\n".join(lines)
+
+
+def _stage_argv(stage: str, seed: int, out_path: str) -> list[str]:
+    """构造子进程命令行：在原始参数上追加阶段参数。"""
+    return [
+        sys.executable,
+        os.path.abspath(__file__),
+        *_ORIGINAL_ARGV,
+        "--stage", stage,
+        "--stage-seed", str(seed),
+        "--stage-out", out_path,
+    ]
+
+
+def _spawn_stage(stage: str, seed: int) -> dict[str, object]:
+    """在子进程里跑一个「一个环境」的阶段，读回它写的 JSON。
+
+    为什么必须开子进程：IsaacLab 在本配置下一个进程只能创建一个环境。第一个环境
+    建好之后，``ROUGH_TERRAINS_CFG`` 这类模块级单例会被运行时长出自引用，之后
+    任何 cfg 构造（新实例 ``env_cfg_cls()``、开场存的 pristine 深拷贝、重新深拷贝）
+    都会在 ``cfg.validate()`` 里 RecursionError。三种补救方式都实测失败，所以只能
+    一个环境一个进程。
+    """
+    handle = tempfile.NamedTemporaryFile(prefix=f"eval_kit_{stage}_", suffix=".json", delete=False)
+    handle.close()
+    out_path = handle.name
+    argv = _stage_argv(stage, seed, out_path)
+    print(f"[eval_kit] 子进程 stage={stage} seed={seed}", flush=True)
+    completed = subprocess.run(argv)
+    if completed.returncode != 0:
+        os.unlink(out_path)
+        raise SystemExit(f"子进程 stage={stage} seed={seed} 失败（退出码 {completed.returncode}）")
+    try:
+        with open(out_path, encoding="utf-8") as file_handle:
+            return json.load(file_handle)
+    finally:
+        os.unlink(out_path)
+
+
+def _recorder_report(recorder) -> dict[str, object]:
+    return {
+        "feet": recorder.feet_names if recorder else [],
+        "support": recorder.support_names if recorder else [],
+        "joints": recorder.joint_names if recorder else [],
+        "mass_kg": recorder.mass_kg if recorder else None,
+        "warnings": recorder.warnings if recorder else [],
+    }
+
+
+def _run_parent(args_cli: argparse.Namespace) -> None:
+    """父进程：不启动仿真，只做编排与合并。每个环境交给一个子进程。"""
+    import numpy as np
+
+    from eval_kit.core.aggregation import aggregate_runs, build_upper_limit_matrix
+    from eval_kit.core.report import filter_metrics, format_aggregate_summary, format_metric_result
+    from eval_kit.core.report import format_rollout_summary, format_upper_limit_summary
+
+    profile = get_profile(args_cli.profile)
+    protocol = _apply_overrides(get_protocol(args_cli.protocol), args_cli)
+    spec = get_terrain_spec(protocol.terrain_spec)
+    if args_cli.robot_mass is not None:
+        profile = replace(profile, mass_kg=args_cli.robot_mass)
+
+    if not args_cli.checkpoint:
+        raise SystemExit(
+            "编排进程需要 --checkpoint 才能派发子进程（父进程不加载 hydra，"
+            "因此无法用 --load-run 按实验名查找）。请传入 checkpoint 路径。"
+        )
+    resume_path = os.path.abspath(os.path.expanduser(args_cli.checkpoint))
+    if not os.path.isfile(resume_path):
+        raise SystemExit(f"找不到 checkpoint：{resume_path}")
+
+    print("[eval_kit] 协议预设:", args_cli.protocol)
+    print("[eval_kit] profile   :", profile.name)
+    print("[eval_kit] checkpoint:", resume_path)
+    print("[eval_kit] 口径      :", json.dumps(protocol.to_dict(), ensure_ascii=False))
+    print("[eval_kit] 地形规格  :", spec.to_dict())
+    print("[eval_kit] 种子      :", list(protocol.seeds), "（每个种子一个子进程）")
+
+    result: dict[str, object] = {
+        "suite": "eval_kit",
+        "task": args_cli.task,
+        "eval_type": args_cli.eval_type,
+        "checkpoint": resume_path,
+        "profile": profile.to_dict(),
+        "protocol": protocol.to_dict(),
+        "terrain_spec": spec.to_dict(),
+        "seeds": list(protocol.seeds),
+        "runner": "one-env-per-process",
+    }
+
+    wants_upper = args_cli.eval_type in ("upper_limits", "all")
+
+    if wants_upper:
+        terrain_types = _terrain_types_for(args_cli, spec)
+        per_seed: list[dict[str, object]] = []
+        pooled: dict[str, list[float]] = {"success": [], "max_progress": [], "max_tilt": []}
+        pooled_levels: list[np.ndarray] = []
+        pooled_types: list[np.ndarray] = []
+        recorder_report: dict[str, object] = {}
+
+        for seed in protocol.seeds:
+            data = _spawn_stage("upper", seed)
+            recorder_report = data["recorder"]  # type: ignore[assignment]
+            per_env = data["per_env"]  # type: ignore[assignment]
+            level_ids = np.asarray(data["level_ids"])  # type: ignore[arg-type]
+            type_ids = np.asarray(data["type_ids"])  # type: ignore[arg-type]
+            for key in pooled:
+                pooled[key].extend(per_env[key])  # type: ignore[index]
+            pooled_levels.append(level_ids)
+            pooled_types.append(type_ids)
+            matrix = build_upper_limit_matrix(
+                per_env, level_ids, type_ids, terrain_types, spec, protocol.num_levels  # type: ignore[arg-type]
+            )
+            per_seed.append(
+                {
+                    "seed": seed,
+                    "scores": matrix["scores"],
+                    "matrix": matrix["matrix"],
+                    "metrics": data["metrics"],
+                    "early_resets": data["early_resets"],
+                }
+            )
+            if args_cli.verbose_terrain:
+                print(f"[eval_kit] seed {seed}: " + "  ".join(f"{k}={v:.2f}" for k, v in matrix["scores"].items()))
+
+        pooled_matrix = build_upper_limit_matrix(
+            pooled, np.concatenate(pooled_levels), np.concatenate(pooled_types), terrain_types, spec, protocol.num_levels
+        )
+        result["terrain_types"] = terrain_types
+        result["upper_limits"] = pooled_matrix
+        result["upper_limits_per_seed"] = per_seed
+        result["recorder"] = recorder_report
+        print()
+        print(format_upper_limit_summary(pooled_matrix))
+        first_metrics = per_seed[0]["metrics"]  # type: ignore[index]
+        print()
+        print(format_rollout_summary(first_metrics))  # type: ignore[arg-type]
+
+        for entry in per_seed:
+            early = entry["early_resets"]  # type: ignore[index]
+            if early and early["polluted"]:  # type: ignore[index]
+                print()
+                print(_format_early_reset_warning(early, protocol, protocol.minimum_terrain_size_for_bounds()))  # type: ignore[index,arg-type]
+                break
+
+        if args_cli.eval_type == "all" and args_cli.include_summary:
+            runs = []
+            for seed in protocol.seeds:
+                data = _spawn_stage("plain", seed)
+                runs.append((seed, data["metrics"]))  # type: ignore[arg-type]
+            result["summary"] = aggregate_runs(runs)  # type: ignore[arg-type]
+            print()
+            print("[混合地形汇总]")
+            print(format_aggregate_summary(result["summary"]))  # type: ignore[arg-type]
+            if runs:
+                result["summary_metrics"] = filter_metrics(runs[0][1], args_cli.eval_type)
+    else:
+        runs = []
+        recorder_report = {}
+        early_any = None
+        for seed in protocol.seeds:
+            data = _spawn_stage("plain", seed)
+            runs.append((seed, data["metrics"]))  # type: ignore[arg-type]
+            recorder_report = data["recorder"]  # type: ignore[assignment]
+            early_any = data["early_resets"]
+        aggregate = aggregate_runs(runs)  # type: ignore[arg-type]
+        result["aggregate"] = aggregate
+        result["metrics"] = filter_metrics(runs[0][1], args_cli.eval_type)
+        result["recorder"] = recorder_report
+        result["early_resets"] = early_any
+        print()
+        print(format_metric_result(args_cli.eval_type, runs[0][1]))
+        if len(runs) > 1:
+            print()
+            print(format_aggregate_summary(aggregate))
+        if early_any and early_any["polluted"]:  # type: ignore[index]
+            print()
+            print(_format_early_reset_warning(early_any, protocol, protocol.minimum_terrain_size_for_bounds()))  # type: ignore[index,arg-type]
+
+    warnings = result.get("recorder", {}).get("warnings")  # type: ignore[union-attr]
+    if warnings:
+        print("\n[形态解析警告]")
+        for warning in warnings:  # type: ignore[union-attr]
+            print("  -", warning)
+
+    output_path = _resolve_output_path(args_cli, resume_path, args_cli.eval_type)
+    with open(output_path, "w", encoding="utf-8") as handle:
+        json.dump(result, handle, indent=2, ensure_ascii=False)
+    print(f"\n[eval_kit] 结果已保存：{output_path}")
+    if args_cli.print_json:
+        print(json.dumps(result, indent=2, ensure_ascii=False))
+
+
+def _run_child(
+    args_cli: argparse.Namespace,
+    *,
+    viewer_env_index: int | None,
+    free_camera: bool,
+    livestream: int,
+    pause_s: float,
+) -> None:
+    """子进程：启动仿真，只建一个环境，跑完把结果写进 ``--stage-out``。
+
+    ``viewer_env_index`` / ``free_camera`` / ``livestream`` / ``pause_s`` 都由
+    :func:`main` 在 ``AppLauncher`` 之前算好传进来：构造 ``AppLauncher`` 会改写
+    ``headless`` 并把 ``livestream`` 从 namespace 里 pop 掉，之后再读就晚了。
+    """
+    print(f"[eval_kit][child] stage={args_cli.stage} seed={args_cli.stage_seed} 开始", flush=True)
+    import numpy as np
+
+    from isaaclab_rl.rsl_rl import RslRlOnPolicyRunnerCfg
+    from isaaclab_tasks.utils.hydra import hydra_task_config
+
+    import rl_training  # noqa: F401  触发任务注册
+    import rl_training.tasks  # noqa: F401
+
+    from eval_kit.core.aggregation import build_upper_limit_matrix
+    from eval_kit.core.metrics import compute_metrics
+    from eval_kit.isaaclab import env_setup
+    from eval_kit.isaaclab.policy import load_policy, make_env, resolve_checkpoint
+    from eval_kit.isaaclab.recorder import build_recorder, record_rollout, summarize_early_resets
+
+    profile = get_profile(args_cli.profile)
+    # 子进程只跑一个种子：种子的覆盖由父进程通过 --stage-seed 传进来
+    protocol = replace(_apply_overrides(get_protocol(args_cli.protocol), args_cli), seeds=(args_cli.stage_seed,))
+    spec = get_terrain_spec(protocol.terrain_spec)
+    if args_cli.robot_mass is not None:
+        profile = replace(profile, mass_kg=args_cli.robot_mass)
+
+    @hydra_task_config(args_cli.task, args_cli.agent)
+    def _child(env_cfg, agent_cfg: RslRlOnPolicyRunnerCfg):
+        experiment_name = args_cli.experiment_name or agent_cfg.experiment_name
+        log_root_path = os.path.abspath(os.path.join("logs", "rsl_rl", experiment_name))
+        resume_path = resolve_checkpoint(log_root_path, args_cli.load_run, args_cli.checkpoint)
+
+        cfg = copy.deepcopy(env_cfg)
+        if args_cli.stage == "upper":
+            terrain_types = _terrain_types_for(args_cli, spec)
+            env_setup.configure_upper_limit_env(
+                cfg, spec, protocol, terrain_types,
+                seed=args_cli.stage_seed, device=args_cli.device,
+                viewer_env_index=viewer_env_index, free_camera=free_camera,
+            )
+        else:
+            terrain_types = []
+            env_setup.configure_eval_env(
+                cfg, spec, protocol, num_envs=args_cli.num_envs,
+                device=args_cli.device, seed=args_cli.stage_seed,
+                viewer_env_index=viewer_env_index, free_camera=free_camera,
+            )
+
+        if viewer_env_index is not None:
+            num_envs = int(cfg.scene.num_envs)
+            if not (0 <= viewer_env_index < num_envs):
+                raise SystemExit(
+                    f"--view-env {viewer_env_index} 超出范围：本次共 {num_envs} 个环境（编号 0~{num_envs - 1}）。"
+                )
+            if terrain_types:
+                per_type = protocol.num_levels * protocol.envs_per_cell
+                type_index, remainder = divmod(viewer_env_index, per_type)
+                level_index = remainder // protocol.envs_per_cell
+                print(
+                    f"[eval_kit] 视口跟随 env {viewer_env_index} → 地形 {terrain_types[type_index]}"
+                    f"，难度等级 L{level_index + 1}（共 {num_envs} 个环境）",
+                    flush=True,
+                )
+            else:
+                print(f"[eval_kit] 视口跟随 env {viewer_env_index}（共 {num_envs} 个环境）", flush=True)
+
+        env = make_env(args_cli.task, cfg, agent_cfg.clip_actions)
+        recorder = build_recorder(env, profile)
+        try:
+            # 策略跟仿真放在同一张卡上：默认的 --device 也会覆盖 agent 配置里的
+            # device，否则在别人占着 cuda:0 的机器上会莫名其妙 OOM。
+            policy = load_policy(env, agent_cfg, resume_path, device=args_cli.device)
+            if args_cli.record_frames:
+                policy = _with_frame_capture(policy, args_cli.record_frames, int(args_cli.record_every))
+            if args_cli.stage == "upper":
+                level_ids, type_ids = env_setup.assign_upper_limit_origins(env, spec, protocol, terrain_types)
+            if free_camera:
+                _set_free_camera_overview(env)
+            if pause_s > 0:
+                _hold_before_rollout(
+                    env.unwrapped.sim,
+                    pause_s,
+                    start_height_note=(
+                        f"（本次起始高度被抬高了 {protocol.start_z_offset_m:.2f} m，"
+                        f"所以是离地悬着的）"
+                        if protocol.start_z_offset_m > 0
+                        else ""
+                    ),
+                )
+            trace = record_rollout(env, policy, protocol, profile, recorder)
+            if args_cli.record_frames:
+                _report_recorded_frames(args_cli.record_frames, int(args_cli.record_every), protocol.dt)
+            metrics = compute_metrics(
+                trace, profile, protocol, include_per_env=(args_cli.stage == "upper")
+            )
+            early = summarize_early_resets(trace)
+        finally:
+            env.close()
+
+        if args_cli.save_trace:
+            os.makedirs(args_cli.save_trace, exist_ok=True)
+            tag = "upper" if args_cli.stage == "upper" else args_cli.eval_type
+            trace.save(os.path.join(args_cli.save_trace, f"{tag}_seed{args_cli.stage_seed}.npz"))
+
+        payload: dict[str, object] = {
+            "seed": args_cli.stage_seed,
+            "metrics": metrics,
+            "early_resets": early,
+            "recorder": _recorder_report(recorder),
+        }
+        if args_cli.stage == "upper":
+            payload["per_env"] = metrics.pop("_per_env")
+            payload["level_ids"] = level_ids.cpu().numpy().tolist()
+            payload["type_ids"] = type_ids.cpu().numpy().tolist()
+        with open(args_cli.stage_out, "w", encoding="utf-8") as handle:
+            json.dump(payload, handle)
+
+    _child()
+    print("[eval_kit][child] 完成", flush=True)
+
+
+def main() -> None:
+    parser = _build_parser()
+    AppLauncher.add_app_launcher_args(parser)
+    args_cli, hydra_args = parser.parse_known_args()
+
+    # 父进程只做编排：它不建环境，因此也不需要启动仿真。
+    if args_cli.stage == "":
+        if _livestream_value(args_cli) > 0 and not args_cli.headless:
+            print(
+                "[eval_kit] 注意：--livestream 会自动进入 headless 模式（推流由服务端渲染），"
+                "不需要也不应该再传 --headless。",
+                flush=True,
+            )
+        _run_parent(args_cli)
+        return
+
+    # 这三个值必须在 AppLauncher 之前定下来（见 _viewer_env_index 的说明）。
+    livestream = _livestream_value(args_cli)
+    viewer_env_index = _viewer_env_index(args_cli)
+    free_camera = bool(getattr(args_cli, "free_camera", False)) and _display_enabled(args_cli)
+    pause_s = args_cli.pause_s if args_cli.pause_s is not None else (30.0 if livestream > 0 else 0.0)
+    if args_cli.free_camera and not free_camera:
+        print(
+            "[eval_kit] 提示：--free-camera 只在有画面时（本地 GUI 或 --livestream）有效，"
+            "本次是 headless 且未推流，忽略该参数。",
+            flush=True,
+        )
+    if livestream > 0:
+        _allow_client_resize(args_cli)
+
+    # 子进程：启动仿真，跑一个环境，然后退出。
+    sys.argv = [sys.argv[0]] + hydra_args
+    simulation_app = AppLauncher(args_cli).app
+    try:
+        if livestream > 0:
+            _print_livestream_info(livestream)
+        _run_child(
+            args_cli,
+            viewer_env_index=viewer_env_index,
+            free_camera=free_camera,
+            livestream=livestream,
+            pause_s=pause_s,
+        )
+    except BaseException:
+        # 关键：不能走 simulation_app.close()。它会调用 os._exit(0)，把「子进程失败」
+        # 伪装成「成功」，并且丢掉还没 flush 的 stderr —— 于是父进程只看到退出码 0
+        # 和一个空的 JSON，真实报错（例如资产配置里的 SyntaxError）完全看不到。
+        # 先把报错刷出去，再以非零码强制退出，父进程才会把真实错误报上来。
+        import traceback
+
+        traceback.print_exc()
+        for stream in (sys.stdout, sys.stderr):
+            stream.flush()
+        os._exit(1)
+    for stream in (sys.stdout, sys.stderr):
+        stream.flush()
+    simulation_app.close()
+
+
+
+if __name__ == "__main__":
+    main()

scripts/reinforcement_learning/rsl_rl/play.py

@@ -0,0 +1,285 @@
+# Copyright (c) 2025 Deep Robotics
+# SPDX-License-Identifier: BSD 3-Clause
+
+# Copyright (c) 2024-2025 Ziqi Fan
+# SPDX-License-Identifier: Apache-2.0
+
+# Copyright (c) 2024-2025, The Isaac Lab Project Developers.
+# All rights reserved.
+#
+# SPDX-License-Identifier: Apache-2.0
+
+"""Script to play a checkpoint if an RL agent from RSL-RL."""
+
+"""Launch Isaac Sim Simulator first."""
+
+import argparse
+import os
+import sys
+
+PROJECT_ROOT = os.path.abspath(os.path.join(os.path.dirname(__file__), "..", "..", ".."))
+PROJECT_SOURCE_PATH = os.path.join(PROJECT_ROOT, "source", "rl_training")
+if PROJECT_SOURCE_PATH not in sys.path:
+    sys.path.insert(0, PROJECT_SOURCE_PATH)
+
+from isaaclab.app import AppLauncher
+
+# local imports
+sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__), "..")))
+import cli_args
+
+# add argparse arguments
+parser = argparse.ArgumentParser(description="Train an RL agent with RSL-RL.")
+parser.add_argument("--video", action="store_true", default=False, help="Record videos during training.")
+parser.add_argument("--video_length", type=int, default=200, help="Length of the recorded video (in steps).")
+parser.add_argument(
+    "--disable_fabric", action="store_true", default=False, help="Disable fabric and use USD I/O operations."
+)
+parser.add_argument("--num_envs", type=int, default=None, help="Number of environments to simulate.")
+parser.add_argument("--task", type=str, default=None, help="Name of the task.")
+parser.add_argument(
+    "--agent", type=str, default="rsl_rl_cfg_entry_point", help="Name of the RL agent configuration entry point."
+)
+parser.add_argument("--seed", type=int, default=None, help="Seed used for the environment")
+parser.add_argument(
+    "--use_pretrained_checkpoint",
+    action="store_true",
+    help="Use the pre-trained checkpoint from Nucleus.",
+)
+parser.add_argument("--real-time", action="store_true", default=False, help="Run in real-time, if possible.")
+parser.add_argument("--keyboard", action="store_true", default=False, help="Whether to use keyboard.")
+parser.add_argument(
+    "--visual-only",
+    action="store_true",
+    default=False,
+    help="Display the environment with zero actions without loading an RL checkpoint.",
+)
+# append RSL-RL cli arguments
+cli_args.add_rsl_rl_args(parser)
+# append AppLauncher cli args
+AppLauncher.add_app_launcher_args(parser)
+# parse the arguments
+args_cli, hydra_args = parser.parse_known_args()
+# always enable cameras to record video
+if args_cli.video:
+    args_cli.enable_cameras = True
+
+# clear out sys.argv for Hydra
+sys.argv = [sys.argv[0]] + hydra_args
+
+# launch omniverse app
+app_launcher = AppLauncher(args_cli)
+simulation_app = app_launcher.app
+
+"""Rest everything follows."""
+
+import gymnasium as gym
+import time
+import torch
+
+from rsl_rl.runners import OnPolicyRunner
+
+from isaaclab.devices import Se2Keyboard, Se2KeyboardCfg
+from isaaclab.envs import (
+    DirectMARLEnv,
+    DirectMARLEnvCfg,
+    DirectRLEnvCfg,
+    ManagerBasedRLEnvCfg,
+    multi_agent_to_single_agent,
+)
+from isaaclab.managers import ObservationTermCfg as ObsTerm
+from isaaclab.utils.assets import retrieve_file_path
+from isaaclab.utils.dict import print_dict
+try:
+    from isaaclab.utils.pretrained_checkpoint import get_published_pretrained_checkpoint
+except ModuleNotFoundError:
+    # Isaac Lab 2.3.x and earlier keep this helper under isaaclab_rl.
+    from isaaclab_rl.utils.pretrained_checkpoint import get_published_pretrained_checkpoint
+from isaaclab_rl.rsl_rl import RslRlOnPolicyRunnerCfg, RslRlVecEnvWrapper, export_policy_as_jit, export_policy_as_onnx
+from isaaclab_tasks.utils import get_checkpoint_path
+from isaaclab_tasks.utils.hydra import hydra_task_config
+
+import rl_training
+import rl_training.tasks  # noqa: F401
+
+from rl_utils import camera_follow
+
+print(f"[INFO] Using rl_training package from: {rl_training.__file__}")
+
+
+@hydra_task_config(args_cli.task, args_cli.agent)
+def main(env_cfg: ManagerBasedRLEnvCfg | DirectRLEnvCfg | DirectMARLEnvCfg, agent_cfg: RslRlOnPolicyRunnerCfg):
+    """Play with RSL-RL agent."""
+    task_name = args_cli.task.split(":")[-1]
+    # override configurations with non-hydra CLI arguments
+    agent_cfg = cli_args.update_rsl_rl_cfg(agent_cfg, args_cli)
+    env_cfg.scene.num_envs = args_cli.num_envs if args_cli.num_envs is not None else 50
+
+    # set the environment seed
+    # note: certain randomizations occur in the environment initialization so we set the seed here
+    env_cfg.seed = agent_cfg.seed
+    env_cfg.sim.device = args_cli.device if args_cli.device is not None else env_cfg.sim.device
+
+    # spawn the robot randomly in the grid (instead of their terrain levels)
+    env_cfg.scene.terrain.max_init_terrain_level = None
+    # reduce the number of terrains to save memory
+    if env_cfg.scene.terrain.terrain_generator is not None:
+        env_cfg.scene.terrain.terrain_generator.num_rows = 5
+        env_cfg.scene.terrain.terrain_generator.num_cols = 5
+        env_cfg.scene.terrain.terrain_generator.curriculum = False
+
+    # disable randomization for play
+    env_cfg.observations.policy.enable_corruption = False
+    # Remove perturbations and parameter randomization for reproducible playback.
+    for event_name in (
+        "randomize_rigid_body_material",
+        "randomize_wheel_material",
+        "randomize_support_material",
+        "randomize_rigid_body_mass",
+        "randomize_rigid_body_mass_base",
+        "randomize_rigid_body_inertia",
+        "randomize_com_positions",
+        "randomize_apply_external_force_torque",
+        "randomize_actuator_gains",
+        "randomize_push_robot",
+    ):
+        if hasattr(env_cfg.events, event_name):
+            setattr(env_cfg.events, event_name, None)
+    env_cfg.curriculum.command_levels = None
+
+    if args_cli.keyboard:
+        env_cfg.scene.num_envs = 1
+        env_cfg.terminations.time_out = None
+        env_cfg.commands.base_velocity.debug_vis = False
+        config = Se2KeyboardCfg(
+            v_x_sensitivity=env_cfg.commands.base_velocity.ranges.lin_vel_x[1]/2,
+            v_y_sensitivity=env_cfg.commands.base_velocity.ranges.lin_vel_y[1],
+            omega_z_sensitivity=env_cfg.commands.base_velocity.ranges.ang_vel_z[1],
+        )
+        controller = Se2Keyboard(config)
+        env_cfg.observations.policy.velocity_commands = ObsTerm(
+            func=lambda env: torch.tensor(controller.advance(), dtype=torch.float32).unsqueeze(0).to(env.device),
+        )
+
+    # specify directory for logging experiments
+    log_root_path = os.path.join("logs", "rsl_rl", agent_cfg.experiment_name)
+    log_root_path = os.path.abspath(log_root_path)
+    if args_cli.visual_only:
+        resume_path = None
+        log_dir = os.path.join(log_root_path, "visual_only")
+        print("[INFO] Visual-only mode: no policy checkpoint will be loaded.")
+    else:
+        print(f"[INFO] Loading experiment from directory: {log_root_path}")
+        if args_cli.use_pretrained_checkpoint:
+            resume_path = get_published_pretrained_checkpoint("rsl_rl", task_name)
+            if not resume_path:
+                print("[INFO] Unfortunately a pre-trained checkpoint is currently unavailable for this task.")
+                return
+        elif args_cli.load_run is not None:
+            resume_path = get_checkpoint_path(log_root_path, agent_cfg.load_run, agent_cfg.load_checkpoint)
+        elif args_cli.checkpoint:
+            resume_path = retrieve_file_path(args_cli.checkpoint)
+        else:
+            resume_path = get_checkpoint_path(log_root_path, agent_cfg.load_run, agent_cfg.load_checkpoint)
+        log_dir = os.path.dirname(resume_path)
+
+    # create isaac environment
+    env = gym.make(args_cli.task, cfg=env_cfg, render_mode="rgb_array" if args_cli.video else None)
+
+    # convert to single-agent instance if required by the RL algorithm
+    if isinstance(env.unwrapped, DirectMARLEnv):
+        env = multi_agent_to_single_agent(env)
+
+    # wrap for video recording
+    if args_cli.video:
+        video_kwargs = {
+            "video_folder": os.path.join(log_dir, "videos", "play"),
+            "step_trigger": lambda step: step == 0,
+            "video_length": args_cli.video_length,
+            "disable_logger": True,
+        }
+        print("[INFO] Recording videos during training.")
+        print_dict(video_kwargs, nesting=4)
+        env = gym.wrappers.RecordVideo(env, **video_kwargs)
+
+    # wrap around environment for rsl-rl
+    env = RslRlVecEnvWrapper(env, clip_actions=agent_cfg.clip_actions)
+
+    if args_cli.visual_only:
+        policy = None
+    else:
+        print(f"[INFO]: Loading model checkpoint from: {resume_path}")
+        # load previously trained model
+        ppo_runner = OnPolicyRunner(env, agent_cfg.to_dict(), log_dir=None, device=agent_cfg.device)
+        ppo_runner.load(resume_path)
+
+        # obtain the trained policy for inference
+        policy = ppo_runner.get_inference_policy(device=env.unwrapped.device)
+
+        # extract the neural network module
+        # we do this in a try-except to maintain backwards compatibility.
+        try:
+            # version 2.3 onwards
+            policy_nn = ppo_runner.alg.policy
+        except AttributeError:
+            # version 2.2 and below
+            policy_nn = ppo_runner.alg.actor_critic
+
+        # export policy to onnx/jit
+        export_model_dir = os.path.join(os.path.dirname(resume_path), "exported")
+        export_policy_as_onnx(
+            policy=policy_nn,
+            normalizer=None,
+            path=export_model_dir,
+            filename="policy.onnx",
+        )
+        export_policy_as_jit(
+            policy=policy_nn,
+            normalizer=None,
+            path=export_model_dir,
+            filename="policy.pt",
+        )
+
+    dt = env.unwrapped.step_dt
+    # print(dt, "dt")
+    # reset environment
+    
+    obs = env.get_observations()
+    
+    timestep = 0
+    # simulate environment
+    while simulation_app.is_running():
+        start_time = time.time()
+        # run everything in inference mode
+        with torch.inference_mode():
+            # agent stepping
+            if policy is None:
+                actions = torch.zeros((env.num_envs, env.num_actions), device=env.unwrapped.device)
+            else:
+                actions = policy(obs)
+
+            # env stepping
+            obs, _, _, _ = env.step(actions)
+        if args_cli.video:
+            timestep += 1
+            # Exit the play loop after recording one video
+            if timestep == args_cli.video_length:
+                break
+
+        if args_cli.keyboard:
+            camera_follow(env)
+
+        # time delay for real-time evaluation
+        sleep_time = dt - (time.time() - start_time)
+        if args_cli.real_time and sleep_time > 0:
+            time.sleep(sleep_time)
+
+    # close the simulator
+    env.close()
+
+
+if __name__ == "__main__":
+    # run the main function
+    main()
+    # close sim app
+    simulation_app.close()

scripts/reinforcement_learning/rsl_rl/train.py

@@ -0,0 +1,214 @@
+# Copyright (c) 2025 Deep Robotics
+# SPDX-License-Identifier: BSD 3-Clause
+
+# Copyright (c) 2024-2025 Ziqi Fan
+# SPDX-License-Identifier: Apache-2.0
+
+# Copyright (c) 2024-2025, The Isaac Lab Project Developers.
+# All rights reserved.
+#
+# SPDX-License-Identifier: Apache-2.0
+
+"""Script to train RL agent with RSL-RL."""
+
+"""Launch Isaac Sim Simulator first."""
+
+import argparse
+import os
+import sys
+
+PROJECT_ROOT = os.path.abspath(os.path.join(os.path.dirname(__file__), "..", "..", ".."))
+PROJECT_SOURCE_PATH = os.path.join(PROJECT_ROOT, "source", "rl_training")
+if PROJECT_SOURCE_PATH not in sys.path:
+    sys.path.insert(0, PROJECT_SOURCE_PATH)
+
+from isaaclab.app import AppLauncher
+
+# local imports
+sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__), "..")))
+import cli_args
+
+# add argparse arguments
+parser = argparse.ArgumentParser(description="Train an RL agent with RSL-RL.")
+parser.add_argument("--video", action="store_true", default=False, help="Record videos during training.")
+parser.add_argument("--video_length", type=int, default=200, help="Length of the recorded video (in steps).")
+parser.add_argument("--video_interval", type=int, default=2000, help="Interval between video recordings (in steps).")
+parser.add_argument("--num_envs", type=int, default=None, help="Number of environments to simulate.")
+parser.add_argument("--task", type=str, default=None, help="Name of the task.")
+parser.add_argument(
+    "--agent", type=str, default="rsl_rl_cfg_entry_point", help="Name of the RL agent configuration entry point."
+)
+parser.add_argument("--seed", type=int, default=None, help="Seed used for the environment")
+parser.add_argument("--max_iterations", type=int, default=None, help="RL Policy training iterations.")
+parser.add_argument(
+    "--distributed", action="store_true", default=False, help="Run training with multiple GPUs or nodes."
+)
+# append RSL-RL cli arguments
+cli_args.add_rsl_rl_args(parser)
+# append AppLauncher cli args
+AppLauncher.add_app_launcher_args(parser)
+args_cli, hydra_args = parser.parse_known_args()
+
+# always enable cameras to record video
+if args_cli.video:
+    args_cli.enable_cameras = True
+
+# clear out sys.argv for Hydra
+sys.argv = [sys.argv[0]] + hydra_args
+
+# launch omniverse app
+app_launcher = AppLauncher(args_cli)
+simulation_app = app_launcher.app
+
+"""Rest everything follows."""
+
+import gymnasium as gym
+import os
+import torch
+from datetime import datetime
+
+from rsl_rl.runners import OnPolicyRunner
+
+from isaaclab.envs import (
+    DirectMARLEnv,
+    DirectMARLEnvCfg,
+    DirectRLEnvCfg,
+    ManagerBasedRLEnvCfg,
+    multi_agent_to_single_agent,
+)
+from isaaclab.utils.dict import print_dict
+from isaaclab.utils.io import dump_yaml
+from isaaclab_rl.rsl_rl import RslRlOnPolicyRunnerCfg, RslRlVecEnvWrapper
+from isaaclab_tasks.utils import get_checkpoint_path
+from isaaclab_tasks.utils.hydra import hydra_task_config
+
+import rl_training
+import rl_training.tasks  # noqa: F401
+
+print(f"[INFO] Using rl_training package from: {rl_training.__file__}")
+
+torch.backends.cuda.matmul.allow_tf32 = True
+torch.backends.cudnn.allow_tf32 = True
+torch.backends.cudnn.deterministic = False
+torch.backends.cudnn.benchmark = False
+
+
+def resolve_resume_path(log_root_path: str, load_run: str | None, load_checkpoint: str | None) -> str:
+    """Resolve a resume checkpoint, allowing explicit paths across experiment folders."""
+    if load_checkpoint is not None:
+        checkpoint_path = os.path.expanduser(load_checkpoint)
+        if os.path.isabs(checkpoint_path) or os.path.dirname(checkpoint_path):
+            checkpoint_path = os.path.abspath(checkpoint_path)
+            if not os.path.isfile(checkpoint_path):
+                raise FileNotFoundError(f"Unable to find the checkpoint file: {checkpoint_path}")
+            return checkpoint_path
+
+    if load_run is not None:
+        run_path = os.path.expanduser(load_run)
+        if os.path.isabs(run_path) or os.path.dirname(run_path):
+            run_path = os.path.abspath(run_path)
+            if not os.path.isdir(run_path):
+                raise FileNotFoundError(f"Unable to find the run directory: {run_path}")
+            return get_checkpoint_path(os.path.dirname(run_path), os.path.basename(run_path), load_checkpoint)
+
+    return get_checkpoint_path(log_root_path, load_run, load_checkpoint)
+
+
+@hydra_task_config(args_cli.task, args_cli.agent)
+def main(env_cfg: ManagerBasedRLEnvCfg | DirectRLEnvCfg | DirectMARLEnvCfg, agent_cfg: RslRlOnPolicyRunnerCfg):
+    """Train with RSL-RL agent."""
+    # override configurations with non-hydra CLI arguments
+    agent_cfg = cli_args.update_rsl_rl_cfg(agent_cfg, args_cli)
+    env_cfg.scene.num_envs = args_cli.num_envs if args_cli.num_envs is not None else env_cfg.scene.num_envs
+    agent_cfg.max_iterations = (
+        args_cli.max_iterations if args_cli.max_iterations is not None else agent_cfg.max_iterations
+    )
+
+    # set the environment seed
+    # note: certain randomizations occur in the environment initialization so we set the seed here
+    env_cfg.seed = agent_cfg.seed
+    env_cfg.sim.device = args_cli.device if args_cli.device is not None else env_cfg.sim.device
+
+    for reward_name in (
+        "track_diff_wheel_yaw_exp",
+        "lin_vel_xy_yaw_command_l2",
+        "hipx_joint_pos_penalty",
+        "hipy_joint_pos_penalty",
+    ):
+        reward_term = getattr(env_cfg.rewards, reward_name, None)
+        reward_weight = None if reward_term is None else reward_term.weight
+        print(f"[INFO] Reward term {reward_name}: weight={reward_weight}")
+
+    # multi-gpu training configuration
+    if args_cli.distributed:
+        env_cfg.sim.device = f"cuda:{app_launcher.local_rank}"
+        agent_cfg.device = f"cuda:{app_launcher.local_rank}"
+
+        # set seed to have diversity in different threads
+        seed = agent_cfg.seed + app_launcher.local_rank
+        env_cfg.seed = seed
+        agent_cfg.seed = seed
+
+    # specify directory for logging experiments
+    log_root_path = os.path.join("logs", "rsl_rl", agent_cfg.experiment_name)
+    log_root_path = os.path.abspath(log_root_path)
+    print(f"[INFO] Logging experiment in directory: {log_root_path}")
+    # specify directory for logging runs: {time-stamp}_{run_name}
+    log_dir = datetime.now().strftime("%Y-%m-%d_%H-%M-%S")
+    # The Ray Tune workflow extracts experiment name using the logging line below, hence, do not change it (see PR #2346, comment-2819298849)
+    print(f"Exact experiment name requested from command line: {log_dir}")
+    if agent_cfg.run_name:
+        log_dir += f"_{agent_cfg.run_name}"
+    log_dir = os.path.join(log_root_path, log_dir)
+
+    # create isaac environment
+    env = gym.make(args_cli.task, cfg=env_cfg, render_mode="rgb_array" if args_cli.video else None)
+
+    # convert to single-agent instance if required by the RL algorithm
+    if isinstance(env.unwrapped, DirectMARLEnv):
+        env = multi_agent_to_single_agent(env)
+
+    # save resume path before creating a new log_dir
+    if agent_cfg.resume or agent_cfg.algorithm.class_name == "Distillation":
+        resume_path = resolve_resume_path(log_root_path, agent_cfg.load_run, agent_cfg.load_checkpoint)
+
+    # wrap for video recording
+    if args_cli.video:
+        video_kwargs = {
+            "video_folder": os.path.join(log_dir, "videos", "train"),
+            "step_trigger": lambda step: step % args_cli.video_interval == 0,
+            "video_length": args_cli.video_length,
+            "disable_logger": True,
+        }
+        print("[INFO] Recording videos during training.")
+        print_dict(video_kwargs, nesting=4)
+        env = gym.wrappers.RecordVideo(env, **video_kwargs)
+
+    # wrap around environment for rsl-rl
+    env = RslRlVecEnvWrapper(env, clip_actions=agent_cfg.clip_actions)
+
+    # create runner from rsl-rl
+    runner = OnPolicyRunner(env, agent_cfg.to_dict(), log_dir=log_dir, device=agent_cfg.device)
+    # write git state to logs
+    runner.add_git_repo_to_log(__file__)
+    # load the checkpoint
+    if agent_cfg.resume or agent_cfg.algorithm.class_name == "Distillation":
+        print(f"[INFO]: Loading model checkpoint from: {resume_path}")
+        # load previously trained model
+        runner.load(resume_path)
+
+    # dump the configuration into log-directory
+    dump_yaml(os.path.join(log_dir, "params", "env.yaml"), env_cfg)
+    dump_yaml(os.path.join(log_dir, "params", "agent.yaml"), agent_cfg)
+    # run training
+    runner.learn(num_learning_iterations=agent_cfg.max_iterations, init_at_random_ep_len=True)
+
+    # close the simulator
+    env.close()
+
+
+if __name__ == "__main__":
+    # run the main function
+    main()
+    # close sim app
+    simulation_app.close()