# 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
from rl_utils import camera_follow

# 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
from isaaclab.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

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()