Item	Deployment value
Robot/DDS joint order	FL knee/wheel, FR knee/wheel, HL knee/wheel, HR knee/wheel
Policy joint order	FL/FR/HL/HR knees, then FL/FR/HL/HR wheels
Actor	63 observations, 8 actions, 50 Hz
Low-level loop	200 Hz; MuJoCo step 1 kHz
Knee control	position action scale 0.25, `kp=80`, `kd=2`, limit 75 Nm
Wheel control	velocity action scale 5.0, `kp=0`, `kd=0.6`, limit 28.68 Nm
Wheel	radius 0.096 m, width 0.083 m, mass 1.34 kg
Knee module	mass 0.612 kg
Base	mass 15.882 kg

The actor observation is reconstructed in the exact training order: angular velocity, projected gravity, command, relative joint position, joint velocity, last action, wheel/support contact states, wheel/support force magnitudes, wheel-slip proxy and applied torque. The simulator sends base velocity and contact data through the unused entries 8-15 of the fixed 16-slot DDS joint message. This extension is for sim2sim; sim2real requires equivalent estimators/sensors.

The MuJoCo wheel geoms have their own friction values. Contact restitution is effectively zero. --randomize samples wheel friction, masses, inertias, base COM and control gains once from the training-time ranges.

Policy export

The repository already contains policy/fire_ant_policy.onnx, exported from:

rl_training/logs/rsl_rl/deeprobotics_m20_support/
2026-09-12_17-52-15_stage2_multiterrain/model_39899.pt

Re-export after selecting a different checkpoint:

cd /home/ycpy/huangweifeng/hwf_workspace/rl_training_huoyi/M20_sdk_deploy
python3 scripts/export_fire_ant_policy.py \
  --checkpoint /absolute/path/to/model_xxx.pt \
  --output policy/fire_ant_policy.onnx

The exporter checks the ONNX graph and compares its output against PyTorch.

Build

Install the simulator dependency in the Python environment that has ROS 2 rclpy and drdds:

python3 -m pip install "numpy<2" mujoco

Then build in a ROS 2 Humble workspace. The drdds package must be built or sourced first.

cd /home/ycpy/huangweifeng/hwf_workspace/rl_training_huoyi
source /opt/ros/humble/setup.bash
source /path/to/the/workspace/containing/drdds/install/setup.bash
colcon build --packages-select m20_sdk_deploy --cmake-args -DBUILD_PLATFORM=x86
source install/setup.bash

FireAnt sim2sim

Use the same ROS domain in both terminals. Start MuJoCo first so the controller immediately receives valid joint and IMU data.

Terminal 1 — deterministic flat-ground sanity check:

export ROS_DOMAIN_ID=1
source /opt/ros/humble/setup.bash
source /path/to/the/workspace/containing/drdds/install/setup.bash
python3 /home/ycpy/huangweifeng/hwf_workspace/rl_training_huoyi/M20_sdk_deploy/interface/robot/simulation/mujoco_simulation_ros2.py \
  --robot fire_ant --terrain flat

Terminal 2 — controller and Stage-2 policy:

export ROS_DOMAIN_ID=1
source /opt/ros/humble/setup.bash
source /path/to/the/workspace/containing/drdds/install/setup.bash
source /home/ycpy/huangweifeng/hwf_workspace/rl_training_huoyi/install/setup.bash
ros2 run m20_sdk_deploy rl_deploy

Keyboard sequence:

Press z and wait about 1.5 seconds for the default posture.
Press c to enter RL control.
Hold w to move forward; s requests backward motion and q/e turn.
Press x to lower the robot or r for joint damping.

Stage-2 was trained with zero lateral command, so a/d are intentionally disabled for FireAnt. The policy's Stage-2 forward command range is 0.25-1.2 m/s; the keyboard uses 1.2 m/s while held. With no movement key held, a deployment deadman override holds the default knee pose and commands zero wheel speed, avoiding the zero-command creep caused by Stage-2's positive-only command sampling.

After the flat test, run the mixed 8/12/16 cm obstacle course:

python3 /home/ycpy/huangweifeng/hwf_workspace/rl_training_huoyi/M20_sdk_deploy/interface/robot/simulation/mujoco_simulation_ros2.py \
  --robot fire_ant --terrain rough

Robustness check with one reproducible domain-randomization sample:

python3 /home/ycpy/huangweifeng/hwf_workspace/rl_training_huoyi/M20_sdk_deploy/interface/robot/simulation/mujoco_simulation_ros2.py \
  --robot fire_ant --terrain rough --randomize --seed 1

Add --headless for unattended runs and --realtime-factor 0 to run as fast as possible.

Original M20 compatibility

The old model, interface and 57-to-16 ONNX policy remain available:

# Simulator
python3 interface/robot/simulation/mujoco_simulation_ros2.py --robot m20 --terrain rough

# Controller
ros2 run m20_sdk_deploy rl_deploy --robot=m20