Isaac Lab training

GPU-parallel RL through the isaaclab trainer, with Isaac Lab in its own venv.

The isaaclab trainer runs python -m isaaclab train in a separate Isaac Lab venv and reads its log; strands-robots never imports it (pins conflict; Kit exits the closing process).

uv venv --python 3.12 ~/il && uv pip install --python ~/il/bin/python --prerelease=allow \
  --index https://pypi.nvidia.com --index-strategy unsafe-best-match "isaaclab[rsl-rl,isaacsim]==3.0.0rc1"
export ISAACLAB_PYTHON=~/il/bin/python OMNI_KIT_ACCEPT_EULA=YES

train_policy(action="train", provider="isaaclab", steps=50, output_dir="runs",
             extra={"task": "Isaac-Cartpole", "num_envs": 4096, "physics": "newton_mjwarp", "timeout_s": 600})

It returns a job_id; action="status" reports rewards, success_rate, a failure's cause and checkpoint_dir; action="stop" ends it; action="play" records video; action="record" writes a LeRobotDataset (extra["dataset_dir"]); action="export" writes MLP (not CNN/recurrent) rsl_rl actors for create_policy("rl", checkpoint_dir=...). extra['rl_library'] is rsl_rl, skrl (AMP, multi-agent) or rl_games (Factory, Forge, AutoMate). Your own task package trains once the operator sets STRANDS_ISAACLAB_TASK_PACKAGES=module:register_fn (importable in that venv).

extra['overrides'] (env.*/agent.*, checked against the task config), agent, device, video, deterministic, base_model, resume reach Isaac Lab; learning_rate pins schedule=fixed.

One L40S, 4096 envs: Cartpole 291k steps/s, G1 110k.

Trained this way

Unitree G1, Isaac-Velocity-Rough-G1 on isaacsim_physx: success 99 percent, mean reward 18.5 after 1,500 iterations; 12 recorded episodes, 50 fps (strands-isaaclab-g1-rough-policy)

Unitree H1, Isaac-Velocity-Rough-H1 on isaacsim_physx: success 100 percent, mean reward 25.4 after 1,500 iterations; 12 recorded episodes, 50 fps (strands-isaaclab-h1-rough-policy)

Unitree Go2, Isaac-Velocity-Rough-UnitreeGo2 on isaacsim_physx: success 98 percent, mean reward 30.6 after 1,500 iterations; 12 recorded episodes, 50 fps (strands-isaaclab-go2-rough-policy)

ANYmal D, Isaac-Velocity-Rough-AnymalD on isaacsim_physx: success 90 percent, mean reward 14.4 after 1,500 iterations; 12 recorded episodes, 50 fps (strands-isaaclab-anymald-rough-policy)

Cassie, Isaac-Velocity-Rough-Cassie on isaacsim_physx: success 100 percent, mean reward 30.4 after 1,500 iterations; 12 recorded episodes, 50 fps (strands-isaaclab-cassie-rough-policy)

strands-isaaclab-zoo on the Hub.

Caveats: Isaac Lab 3.0 is an RC; first RTX use compiles shaders (~4 min); PhysX and Newton differ, even in joint order, so export records a deploy_contract create_policy("rl") applies by name.

Edit page