yanshi-berkeley-humanoid_lite-humanoid

Trained policies for berkeley/humanoid_lite/humanoid, published by yanshi-rl-lab.

Each subdirectory is one task. ⛔ A tag snapshots the whole repository; only the subdirectory named below is that release.

# reproduce a release's numbers (pure CPU, no GPU needed)
python scripts/sim2sim/run_gates.py \
    --gates benchmark/gates/<exam>.yaml \
    --hf yanshi-robotics/yanshi-berkeley-humanoid_lite-humanoid@<tag> --subdir <task>

velocity-flat

• Tag: yanshi-bhl-flat-s42-a3
• Exam: velocity-flat/bhl
• Gates cleared: 2 of 4
• Seeds: 1 · Trust: verified
• Code: 9db7285
• Recipe: recipes/berkeley-humanoid_lite-humanoid/velocity-flat.yaml

Reproduce:

python scripts/sim2sim/run_gates.py --gates benchmark/gates/velocity-flat/bhl.yaml --hf jeffliulab/yanshi-berkeley-humanoid_lite-humanoid@bhl-flat-s42-a3 --subdir velocity-flat

Attempt 3 (official HybridRobotics humanoid recipe transcribed verbatim); gates 2/4, no veto. It walks -- straight walk 2.75 m clears its 2.08 m line -- but both slow-command gates fail: the slow turn reaches 7.1 deg against a 45.8 deg line and the slow walk 0.02 m against 0.86 m. Diagnosis on record: the training command envelope (vx +-1.0, heading mode) is far wider than the exam points, so near-zero commands sit in a region the policy barely saw. Listed rather than withheld: a leaderboard that shows only passes reports by omission. First measured 2026-08-02 at commit b79df40; re-measured 2026-08-03 at commit 9db7285 with the weights pulled from the published Hugging Face release and no token, giving the same numbers to every digit. The release publishes the run's policy-9999.onnx under the layout's policy.onnx name; the two are byte-identical. Single seed.

Licence and scope

Weights are MIT, like the code. They contain no vendor model files: robot assets are fetched from their own upstreams under their own licences.

⛔ These policies have never run on physical hardware. They are validated in simulation only, and a second physics engine at that. Any real-robot use is at your own risk.

Generated by scripts/tools/publish_release.py -- do not edit by hand.