Downloads · 30 days
0
unitree-g1-velocity-v28
haixuantao/unitree-g1-velocity-v28
unitree-g1-velocity-v28 is a reinforcement learning model from haixuantao. Use it for the reinforcement learning task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.
RL locomotion policy for the Unitree G1, trained on the zealot trainer (nexus GPU physics, native CUDA). Full 29-DOF body in sim (wrist pitch/yaw welded): the policy commands the 12 leg joints; waist and arms are PD-h…
Access
Public
Updated Aug 6, 2026
Repo size
10.7 MB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors7.3 MB · 68%
At a glance
- Task
- Reinforcement Learning
- License
- apache-2.0
- Access
- Public
- Created
- Aug 5, 2026
- Updated
- Aug 6, 2026
- SHA
- 9571d2de
Model card
From the Hugging Face model README
Unitree G1 (29-DOF) — Velocity Tracking Policy (v28)
RL locomotion policy for the Unitree G1, trained on the zealot trainer (nexus GPU physics, native CUDA). Full 29-DOF body in sim (wrist pitch/yaw welded): the policy commands the 12 leg joints; waist and arms are PD-held, with recorded arm-motion playback during training so the policy tolerates a moving upper body (teleop-friendly).
Snapshots (one folder per training iteration; newest is the default pick):
| snapshot | notes |
|---|---|
velocity_v28_iter44000/ | RECOMMENDED — the run's optimum (3-command MuJoCo sweep, all 20 s survival): smoothest at every command (knee >5 Hz power 17.6/4.4/18.9% at 0.4 / 0.7 / turn) AND fastest at speed (12.25 m on the 0.7 m/s command) |
velocity_v28_iter41500/ | smooth (26% >5 Hz at 0.4 m/s); superseded by iter44000 |
velocity_v28_iter31350/ | fast at 0.4 m/s (8.70 m) but 69% >5 Hz knee chatter; superseded by iter44000 |
velocity_v28_iter12300/ | terrain curriculum at level ~12.8 (26.8% slopes); MuJoCo sim2sim: full 20 s survival, 7.90 m |
velocity_v28_iter7900/ | earlier snapshot at the ~0.10/step reward plateau |
The run targets 50,000 iterations; later snapshots will be added.
Files
velocity_v28_iter7900/policy.onnx— actor network (ONNX, 0.68 MB)- Input:
obs[N, 265]float32 — 5-frame history × 53, obs normalizer (clip ±5) baked into the graph - Output:
action[N, 12]float32 — joint-position offsets;q_target = default_pos + 0.5 × action - Export verified against the float64 numpy reference (max abs err 4.8e-6)
- Input:
velocity_v28_iter7900/policy.safetensors— raw actor checkpointvelocity_v28_iter7900/env.yaml— deploy contract: joint order, PD gains, action scale, default pose, obs spec, DR summary
Key deploy notes
- Ankle gains are kp 40 / kd 2.0 (the v19+ deploy package), NOT the AGILE
table's 20/0.2 — run inference with the gains in
env.yaml. - Control at 50 Hz; obs builder reference:
zealot/examples/biped/sim2sim_xval.py(note its lag-2 last_action convention). - Commands: vx ∈ [−0.8, 0.8] m/s; trained with standing commands, pushes (0.5 m/s / 0.25 rad/s), motor delay 0–4 substeps, terrain slopes to 26.8%.
Training
- Trainer: zealot
biped_train_gpu(PPO, mirror augmentation, obs-normalizer freeze), 4096 envs on an RTX 5090, ~2.5 s/iter - Physics: nexus multibody (native CUDA, sm_120), force-based PD motors, force-sensed foot contact (1 N threshold)
- Robot:
unitree_g1_29dof_agilespec (AGILE actuator parametrization, deploy ankle override)
- Task
- Reinforcement Learning
- License
- apache-2.0
- Created
- Aug 5, 2026
- Updated
- Aug 6, 2026