English | 中文

tron1-rl-deploy-python

Part of the TRON1 open-source ecosystem. All TRON1 repositories and where to start: TRON1 open-source ecosystem.

Python deployment of reinforcement-learning locomotion policies on the TRON1 biped: point-foot (PF_*), sole-foot (SF_*) and wheel-foot (WF_*). main.py loads an ONNX policy and history encoder with onnxruntime and sends joint targets through the LimX low-level SDK, either to the tron1-mujoco-sim simulator or to a real robot.

Highlights

• Ten robot models, one entry point. ROBOT_TYPE selects the model; the PF_ / SF_ / WF_ prefix picks the point-foot, sole-foot or wheel-foot controller.
• Policies from Isaac Gym or Isaac Lab. Every model ships both policy sets; RL_TYPE=isaacgym or RL_TYPE=isaaclab chooses one.
• Sim and robot from the same command. No argument connects to the local simulator; python main.py <robot-ip> drives the robot.
• Plain Python. NumPy, PyYAML, SciPy and ONNX Runtime on the CPU; no ROS, no GPU.

Supported robots

	Family	ROBOT_TYPE	Joints / policy actions
	Point-foot	PF_TRON1A, PF_TRON1B, PF_P441A, PF_P441B, PF_P441C, PF_P441C2	6 / 6
	Sole-foot	SF_TRON1A, SF_TRON1B	8 / 8
	Wheel-foot	WF_TRON1A, WF_TRON1B	8 / 8

Each model directory under controllers/model/ holds params.yaml and the policy pair at policy/$RL_TYPE/policy.onnx and policy/$RL_TYPE/encoder.onnx. The control loop runs at 500 Hz and the policy at 50 Hz (loop_frequency: 500, decimation: 10).

How it works

Quick start

Linux x86_64 or aarch64, Python 3, CPU only. Run the commands below from the directory that holds both clones.

1. Clone this repository and the simulator. Submodules carry the SDK, the robot descriptions and the virtual joystick.

git clone --recurse-submodules https://github.com/limxdynamics/tron1-rl-deploy-python.git
git clone --recurse-submodules https://github.com/limxdynamics/tron1-mujoco-sim.git

2. Install the low-level SDK — the wheel ships in the limxsdk-lowlevel submodule, so initialise it first if you cloned without --recurse-submodules (git submodule update --init):

# Linux x86_64
pip install tron1-rl-deploy-python/limxsdk-lowlevel/python3/amd64/limxsdk-*.whl
# Linux aarch64
pip install tron1-rl-deploy-python/limxsdk-lowlevel/python3/aarch64/limxsdk-*.whl

3. Install the controller dependencies:

pip install numpy pyyaml scipy onnxruntime mujoco

4. Select the model and the policy source. ROBOT_TYPE comes from the table above, RL_TYPE is isaacgym or isaaclab. Both terminals need both variables.

export ROBOT_TYPE=PF_TRON1A
export RL_TYPE=isaacgym

5. Start the simulator, then the controller — one terminal each:

python tron1-mujoco-sim/simulator.py      # terminal 1
python tron1-rl-deploy-python/main.py     # terminal 2

The robot falls in the simulator until the controller is running. Once main.py is up, press Backspace in the MuJoCo window to reset the robot; it stands up and starts walking in place. Drive it with the virtual joystick below.

If pip refuses to install outside a virtual environment (externally-managed-environment), or the simulator needs more setup, see the tron1-mujoco-sim README.

Virtual joystick

./tron1-mujoco-sim/robot-joystick/robot-joystick

Axes, as read by the controllers:

Axis	Command	Range
axes[1]	Forward / backward	±0.5 m/s
axes[0]	Sideways	±0.5 m/s
axes[2]	Turn (yaw)	±0.5 rad/s

Running on the robot

Pass the robot's IP address to main.py, with ROBOT_TYPE and RL_TYPE set as in step 4:

python main.py <robot-ip>

On the robot the controller does not start by itself. It waits until the robot reports calibration state 0 (printed as Calibration state: 0), then starts when you press L1 + Y on the remote. L1 + X stops the controller: it sends zero stiffness with damping on every joint and exits.

Buttons	Action
L1 + Y	Start the controller (after calibration state 0)
L1 + X	Stop: zero stiffness, damping only, then exit
Sticks	Velocity commands, see the joystick section

Verification

Checked in the MuJoCo simulator with tron1-mujoco-sim, limxsdk 3.4.0 and ONNX Runtime 1.31 on Linux x86_64. Every ROBOT_TYPE was run with both RL_TYPE=isaacgym and RL_TYPE=isaaclab and zero joystick command. After the reset, every combination stood up and held its base height for 10 s. Point-foot models step in place and drift 0.7 to 1.7 m over those 10 s; sole-foot and wheel-foot models drift 0.4 m or less. The real-robot path (main.py <robot-ip>) was not tested on hardware.

License

Apache License 2.0, see LICENSE and NOTICE.