Whole-body contact force sensing for Unitree G1 — artifacts
Binary artifacts for stevenryanrobot/wholebodycontact (code, evals and docs live there; this repo holds what git can't: datasets, checkpoints, low-level policies, demo videos). State: milestone W1, 2026-09-08.
Layout
| path | what | size |
|---|---|---|
models/ |
all 21 trained force-sensing checkpoints (.pt) + their training json — deployed ts_p35_upper_12k.pt and every ablation row of evaluation/comparison_ablation_2026-09-08.md |
385 MB |
models/grfnet/ |
learned-GRF nets used by the p35l channel | 80 MB |
policies/ceer/ |
CEER low-level policy (authors: btx0424@SUSTech, Qingzhou Lu) | 54 MB |
policies/sonic/ |
Sonic policy family (encoder/decoder/planner onnx + configs) | 825 MB |
datasets/wb30l/ |
wb30l training + held-out set: 17 h5 files, 30-link labels, controllers A–G (E_vstiff is the frozen held-out controller) | 21 GB |
demos/ |
force-control demo videos (constant-force push, hybrid wipe, admittance; incl. real-collision-wall variants) | ~10 MB |
Quick download
from huggingface_hub import snapshot_download
# everything except the 21 GB dataset:
snapshot_download("stevenryanrobot/wholebodycontact", ignore_patterns=["datasets/*"], local_dir="wbc_artifacts")
# the dataset too (place under data/wbc/mthd2.2/wb30l in the code repo):
snapshot_download("stevenryanrobot/wholebodycontact", allow_patterns=["datasets/wb30l/*"], local_dir="wbc_artifacts")
See REPRODUCING.md in the code repo for where each piece goes on disk.