Training console
Train a humanoid
Pick a humanoid, pick a task, then train it here in the tab on MuJoCo 3.14.0 WebAssembly, or queue it for an NVIDIA GPU worker running Dropbear's stack (Isaac Lab or mjlab, rsl-rl PPO). Both export an ONNX policy with the same obs/action contract, and both replay in the viewport.
- Humanoids
- 28
- With a sim model
- 27
- Trainable in a tab
- 27
1 · Humanoid
unitree/g1 · Unitree Robotics
2 · Where it trains
3 · Task
Downloads 19 MB once (cached), then starts one MuJoCo instance per worker.
What the tab does
- Real PPO (rsl-rl's algorithm and defaults: GAE, clipped surrogate and value loss, adaptive KL learning rate; a smaller 64x64 network and 0.3 initial action noise to fit a CPU) on real MuJoCo 3.14 physics, the same MJCF and commit the robot page simulates.
- Parallel environments in Web Workers, one MuJoCo instance each; gradients are computed on every core and summed.
- Observations, actions and rewards follow the rsl-rl locomotion configs Dropbear trains with: joint positions and velocities, projected gravity, base angular velocity, last action; PD position targets around the standing pose.
- The policy you download replays in @hs/sim's PolicyRunner exactly as the viewport shows it.
What it cannot do
- Scale. With 12 workers a laptop CPU ran 6 to 9 thousand policy steps a second here (G1, H1 and T1 learned to stand in 2 to 5 minutes); a laptop GPU ran 30 thousand, a data-centre GPU runs hundreds of thousands. Walking with velocity commands, rough terrain and motion imitation need 10^8 to 10^9 steps: those are GPU jobs.
- Walking: the stand task learns balance and push recovery, not locomotion.
- Hardware readiness. No motor model, latency, sensor noise or mass randomisation. A tab policy is a simulation result; do not run it on a robot without a hardware-faithful retrain.
- Keep going when the tab sleeps: background tabs throttle, and closing the tab ends the run.
Tab tasks: stand, recover from pushes, step in place. GPU tasks: stand, walk (velocity tracking), push recovery. Motion imitation is listed where a retargeted motion library exists, and offered once the worker generates tracking environments.