English | 中文

_{For partnership and procurement inquiries, contact opensource@limxdynamics.com}

TRON2_YG_LAB

Part of the TRON2 open-source ecosystem. Where to start and the other TRON2 repositories: TRON2 quick start.

Reinforcement learning training stack for the LimX TRON2A biped, built on Isaac Lab with PPO. It trains blind locomotion policies for the SF (sole-foot) and WF (wheel-foot) bases, and for the SFYG / WFYG variants that carry a 6-DoF arm and a two-finger gripper. Built for the ATEC challenge.

Scope: blind (no perception) velocity tracking on flat and rough terrain (no stairs). Trained policies are not included: you train them here and export TorchScript / ONNX.

Highlights

• Four robots, one stack. SF, WF, SFYG and WFYG each register a Flat and a Rough training task plus matching -Play-v0 tasks (16 task IDs).
• Arm as payload. In SFYG / WFYG a dedicated PD group holds the arm and gripper in a fixed pose, so the policy learns to walk with the extra mass without the arm entering the action or observation space.
• Rough terrain with curriculum. Procedural flat / waves / boxes / random-rough terrain on a 10×16 grid, difficulty 0→1.
• Velocity estimator built in. An MLP encoder reads 10 steps of observation history and estimates base velocity for the actor.
• Deployment-ready export. play.py writes TorchScript and ONNX files for the policy and the encoder into exported/ next to the checkpoint.

Model lineup

	Model	Description	Task ID prefix
	SF_TRON2A	Biped with sole feet (ankle pitch).	Isaac-Limx-SF-TRON2A-...
	WF_TRON2A	Biped with wheel feet.	Isaac-Limx-WF-TRON2A-...
	SFYG_TRON2A	Sole-foot base + 6-DoF arm + two-finger gripper (locked).	Isaac-Limx-SFYG-TRON2A-...
	WFYG_TRON2A	Wheel-foot base + 6-DoF arm + two-finger gripper (locked).	Isaac-Limx-WFYG-TRON2A-...

Images are rendered in MuJoCo from the bundled robot_description models in the initial pose used for training (arm at zero, gripper fully open).

Quick start

Look at the robots (no Isaac Sim needed):

git clone --recurse-submodules https://github.com/limxdynamics/TRON2_YG_LAB.git
cd TRON2_YG_LAB
pip install mujoco
python -m mujoco.viewer --mjcf=robot_description/tron2/SFYG_TRON2A/xml/robot.xml

Swap SFYG_TRON2A for SF_TRON2A, WF_TRON2A or WFYG_TRON2A.

Train a policy (Isaac Sim 4.5.0 + Isaac Lab installed, see Installation):

python scripts/rsl_rl/train.py --task Isaac-Limx-SFYG-TRON2A-Blind-Rough-v0 --num_envs 4096 --headless

Repository structure

.
├── exts/bipedal_locomotion/   # Isaac Lab extension: env / asset / MDP / robot configs
├── rsl_rl/                    # Vendored rsl_rl fork (PPO, on-policy runner, MLP encoder); train.py puts it first on sys.path
├── scripts/rsl_rl/            # Entry points train.py / play.py / cli_args.py (scripts, not a package)
├── robot_description/         # Git submodule: URDF / USD / MJCF / meshes for TRON2
└── doc/                       # README images

Installation

Requires Isaac Sim 4.5.0 + Isaac Lab (isaaclab / isaaclab_tasks / isaaclab_rl importable), Python 3.10 and an NVIDIA GPU (≥ 12 GB VRAM recommended for 4096 environments). Clone as in Quick start, then:

git submodule update --init --recursive   # only if cloned without --recurse-submodules
pip install -e exts/bipedal_locomotion
pip install -e rsl_rl

The USD files in the robot_description submodule are loaded when training or play starts; without them, spawning the robot fails.

Training

python scripts/rsl_rl/train.py --task <TASK> --num_envs 4096 --headless

<TASK> is Isaac-Limx-{SF,WF,SFYG,WFYG}-TRON2A-Blind-{Flat,Rough}-v0, registered in robots/init.py. Flat is a plain plane; Rough uses BLIND_ROUGH_TERRAINS_CFG in cfg/SF_TRON2A/terrains_cfg.py and cfg/WF_TRON2A/terrains_cfg.py, which YG variants reuse.

Common options:

• --video --video_interval 24000 --video_length 400 — record videos (auto-enables --enable_cameras)
• --max_iterations N — override the maximum iteration count in the PPO cfg

Logs go to logs/rsl_rl/<experiment_name>/<timestamp>_<run_name>/.

Resume training

Pass --resume True explicitly; agent_cfg.resume defaults to False (train.py:130-139).

python scripts/rsl_rl/train.py --task <TASK> --num_envs 4096 --headless --resume True \
    --checkpoint_path logs/rsl_rl/<experiment_name>/<timestamp>_<run_name>/model_<iter>.pt

Or replace --checkpoint_path with --load_run <run_dir> --checkpoint <model_file>; both accept regex (e.g. --load_run ".*") and pick the latest match under logs/rsl_rl/<experiment_name>/ in lexicographic order.

• Use the same task ID as the original run, or loading fails on obs/action dimensions; changing reward weights is fine. Logs go to a new timestamped directory.
• --max_iterations is a cap, not an increment: trained to 1500 and want 1000 more → --max_iterations 2500.

Play and export

Use the -Play-v0 task ID. Play cfg uses fewer envs, disables domain randomization and simplifies terrain.

python scripts/rsl_rl/play.py --task Isaac-Limx-SF-TRON2A-Blind-Rough-Play-v0 --num_envs 32 \
    --checkpoint_path logs/rsl_rl/sf_tron_2a_rough/<run>/model_<iter>.pt

play.py also writes TorchScript and ONNX (policy.onnx, encoder.onnx) to exported/ next to the checkpoint.

YG Variant Design

The arms remain locked in a fixed pose throughout (arm1~6 = 0 rad, gripper1/2 = 0.05 m), held by a dedicated arm_lock ImplicitActuator group (stiffness 800, damping 40) in the asset config performing PD lock.

• Arm joints are not in joint_order_name, so the policy neither commands nor observes them: the arm is payload only.
• Domain randomization, reset and the dof_limits reward exclude arm joints in the YG env cfg, so they neither disturb the lock PD nor add spurious penalties.

For the full arm exclusion checklist, see exts/bipedal_locomotion/bipedal_locomotion/tasks/locomotion/robots/limx_solefoot_yg_tron2a_env_cfg.py and limx_wheelfoot_yg_tron2a_env_cfg.py.

Task wiring

How a task ID is wired up, using Isaac-Limx-SF-TRON2A-Blind-Flat-v0 as an example:

1. gym.register: tasks/locomotion/robots/__init__.py binds (env_cfg, ppo_cfg) to the task ID
2. Env cfg: tasks/locomotion/robots/limx_solefoot_tron2a_env_cfg.py inherits from tasks/locomotion/cfg/SF_TRON2A/limx_base_env_cfg.py::SF_TRON2A_EnvCfg, attaches assets + modifies MDP
3. MDP terms: tasks/locomotion/mdp/{rewards,events,observations,curriculums,commands}.py
4. PPO cfg: tasks/locomotion/agents/limx_rsl_rl_ppo_cfg.py, using the project's custom RslRlPpoAlgorithmMlpCfg type (not the upstream class)
5. Asset cfg: assets/config/<robot>_cfg.py, spawns USD (from robot_description/) + init joint pos + actuators

To add a new robot variant: increment one copy at each of the 5 layers above without modifying the existing TRON2 training stack.

Submodule: robot_description

The robot models come from tron2-robot-description (URDF / xacro / MuJoCo XML / meshes / USD for the TRON2 variants). This repository pins it at commit 8a11c12, where the models live under tron2/<MODEL>/.

Do not update the submodule to the latest main. Later versions of tron2-robot-description moved the models to tron2a/<MODEL>/, so the asset configs in exts/bipedal_locomotion/bipedal_locomotion/assets/config/ would no longer find the USD files. Update the submodule only together with those paths.

Related repositories

Repository	Description
tron2_rl_lab	Isaac Lab RL for SF / WF and the full-body dual-arm DASF_TRON2A
tron2_mujoco_sim	MuJoCo simulator for sim-to-sim checks
tron2_rl_deploy_python · tron2_rl_deploy_ros	Policy deployment in Python / ROS

License

Apache 2.0.