ufo-g1-policy

UFO G1 FB policy

The G1 FB-CPR-Aux policy used in the October 15, 2026 Unitree G1 hardware session: tracking dance1_subject2, goal reaching, and reward-conditioned forward walking at 0.7 m/s. The ONNX actor, its metadata/configuration and task z files are the same files used in that session.

The source experiment is g1_fb_pbfm_dr_reward_w1024_depth6_shared_bf16_192m_x8_20261011_234503. release_manifest.json records hashes and the training step. This release replaces the older artifacts on main; old files remain available at revision fc5efb7b3bb7a82270abaeabb0cf3c1194f1c7e6. Do not combine the old z files with this new model.

Deploy

Use the UFO deploy branch and its current README:

git clone --branch deploy --single-branch https://github.com/Roboparty/UFO.git UFO-Deploy
cd UFO-Deploy
python -m pip install -r requirements/runtime.txt
python download_hf_model.py

The downloader puts the ONNX, metadata, deployment config and three task z files in model/g1_policy/. The repository task defaults match this release. Custom commands may explicitly select the included task configurations:

python -m sim_env.simulator  # terminal A
python deploy.py --mode sim2sim --task tracking --task-config model/g1_policy/tasks/tracking.yaml
python deploy.py --mode sim2sim --task goal --task-config model/g1_policy/tasks/goal.yaml
python deploy.py --mode sim2sim --task reward --task-config model/g1_policy/tasks/reward.yaml

Run one policy task at a time in terminal B. For onboard control, follow the deploy README. A initializes, R1 arms, B starts tracking/reward, and R2 stops; goal executes on R1. Sim2sim rollout videos are included as previews, not as recordings of the physical test. The backward encoder is included for teleop; this release does not claim a new physical teleop validation.

PyTorch inference

From the current UFO main branch:

uv sync --locked
bash scripts/download_data.sh g1_lafan
uv run hf download xuewang/ufo-g1-policy --local-dir model
./run_inference.sh tracking --model-folder model/g1_policy \
  --data-path ufo/data/lafan_29dof.pkl --motion-list 7 --save-mp4
./run_inference.sh goal --model-folder model/g1_policy \
  --data-path ufo/data/lafan_29dof.pkl --save-mp4

checkpoint/model/ contains the complete saved model weights and architecture, including the backward/forward networks and critics. Optimizer state and the approximately 47 GB rank-0 replay buffer are not included, so this is not an exact training-resume package. Computing new reward z requires a compatible saved replay buffer; the provided reward_inference/reward_locomotion.pkl already contains the forward-0.7 m/s z used on hardware.

config.json is the configuration used for the tested inference run, with machine-local paths made relative to UFO. Obsolete DR/auxiliary-reward fields had been removed for inference; weights and exported actions are unchanged. training_config.json preserves the original training configuration for reference and is not the current inference entrypoint.

Files

g1_policy/ contains checkpoint/model/, config.json, training_config.json, exported/, tracking_inference/, goal_inference/, reward_inference/, tasks/, and release_manifest.json. Always keep the ONNX with its matching .meta.json and .deploy.yaml sidecars.

For RP1 UFO deployment, use Roboparty/roboparty_deploy; this model is G1-specific. UFO is released for non-commercial research under CC BY-NC 4.0.