Unitree G1: carry, walk, doors (toolbot)
Collection
G1 policies trained in mjlab on HF Jobs (~$1 each), measured for energy: walk/stand, carry from PRISM clips. • 3 items • Updated
Part of Ethgar's robot lab notebook: rung T0 of toolbot (a Unitree G1 learning to carry tools): stands still with 2.3 cm drift over 10 s, walks 0.44 m/s, turns, no falls. Trained on HF Jobs (L4, 66 min, ~$0.90). The code repository is private for now.
toolbot G1 velocity policy (Mjlab-Velocity-Flat-Unitree-G1)
mjlab's G1 29-DoF velocity-tracking task (walks, turns, and stands still at a zero command), trained with PPO (rsl_rl) as rung T0 of the toolbot project: the Mac -> HF Jobs -> Hub -> Mac loop.
Evaluate on a Mac: uv run python eval_policy.py --hub-repo <this repo> in toolbot's train/ project.
Run
{
"task": "Mjlab-Velocity-Flat-Unitree-G1",
"mjlab": "1.6.0",
"mujoco": "3.11.0",
"mujoco_warp": "3.11.0",
"torch": "2.14.1",
"num_envs": 4096,
"iterations": 3000,
"final_checkpoint": "model_2999.pt",
"train_seconds": 3967.5,
"seconds_per_iteration": 1.322,
"device": "gpu",
"flavor": "l4x1",
"host": "j-ethgar-6abe2cc7fbc85ba68236134c-cvtsi11z-01585-zkv2f",
"toolbot_commit": "aecfd9f"
}
Reward weights
| term | weight |
|---|---|
| track_linear_velocity | 2.0 |
| track_angular_velocity | 2.0 |
| upright | 1.0 |
| pose | 1.0 |
| body_ang_vel | -0.05 |
| angular_momentum | -0.02 |
| dof_pos_limits | -1.0 |
| action_rate_l2 | -0.1 |
| air_time | 0.0 |
| foot_clearance | -2.0 |
| foot_swing_height | -0.25 |
| foot_slip | -0.1 |
| soft_landing | -1e-05 |
| self_collisions | -1.0 |
Exact configs: params/env.yaml, params/agent.yaml. Curves: tensorboard events.