stevenryanrobot/wholebodycontact

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). Last update: 2026-10-30 (end2end / Sonic contact-estimation line: Sonic v1.1 raw data, cross-controller sets, all final checkpoints) (earlier: Route B AMASS models + datasets, mthd2.0/2.1/2.2 checkpoints, GR00T N1.7 fine-tunes).

Layout

path	what	size
models/	mthd2.3 force-sensing checkpoints (.pt) + training json. Deployed: ts_p35_upper_12k.pt (trained on datasets/wb30l). Since 2026-10-16 also ts_p35_upper_12k_amass{A,AB}[_s1,_s2] (Route B, trained with datasets/wb30_amass) and the matching ts_p35_upper_12k_base_s{1,2} seed controls	~550 MB
models/grfnet/	wb30l p35* checkpoints (learned-GRF channel arms; name kept for old links)	80 MB
models/grfnet_estimator/	the 200 Hz ground-reaction estimator itself (grfnet_{free,all}.npz)	10 MB
models/mthd2.2_wb30/	30-link resultant-wrench models on datasets/wb30 (wb30_s1.pt is used by the contact_chat demo)	162 MB
models/mthd2.0/	K-slot multi-contact deploy checkpoints (deploy_slot-*.pt)	66 MB
models/mthd2.1/	dense force-field checkpoints (f2_field*.pt) + F-series result json	110 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
groot_n17/	GR00T N1.7 → SONIC fine-tunes (HF-format model dirs, final weights only): n17_sonic_apple_v12cl2_srv (apple→plate, 12/20), n17_sonic_tasks457 (box/cart/door, box 7/10), n17_sonic_tasks457_force / _noforce (hand-force estimate as state input vs. no-force control)	4 × 8.9 GB
datasets/wb30l/	wb30l training + held-out set: 17 h5, 30-link labels, controllers A–G (E_vstiff = frozen held-out controller), learned-GRF observer	21 GB
datasets/wb30/	the same 17 shards with the oracle-free p35o observer (frozen manifest evaluation/frozen_wb30_manifest.json)	18 GB
datasets/wb30_amass/	Route B: AMASS clips executed by Sonic in MuJoCo with the wb30 force regimes, tier A (perfect model) / B (DR); heldout/ = clip id % 10 == 0	12 GB
datasets/grfnet/	200 Hz training pairs for the ground-reaction estimator	0.5 GB
datasets/sonic/	end2end line raw Sonic v1.1 collection (q, qd, raw tau, act, IMU, 30-link labels, per-foot GRF): train/{G3,G4v2,G5,G5v2,G6}, thr, test, test_long, test_wb → local data/wbc/end2end/sonic/	178 GB
end2end/xpol/	cross-controller evaluation sets (beyondmimic, ceer, fastsac, twist2, unitree_rl)	40 GB
end2end/models*/, end2end/grf_net/	end2end / residual / CFM final checkpoints	~1.5 GB
end2end/e2net2*/, end2end/preds*/	network-2 records and prediction caches (regenerable; uploaded to save re-running)	~115 GB
models/sonicv2/	sonicv2 G1–G5 final checkpoints (35)	0.7 GB
datasets/mthd1_cross/cross_A_base.h5	reference h5 the collectors read (kp/kd, 24-link body_names)	0.56 GB
demos/	force-control demo videos	~10 MB

The groot_n17/ weights are derivatives of NVIDIA GR00T N1.7 and are governed by NVIDIA's model license, not the MIT license above.

Quick download

from huggingface_hub import snapshot_download
# all small checkpoints + policies (no datasets, no GR00T):
snapshot_download("stevenryanrobot/wholebodycontact", ignore_patterns=["datasets/*", "groot_n17/*"], local_dir="wbc_artifacts")
# one dataset (place under data/wbc/mthd2.2/<name> in the code repo):
snapshot_download("stevenryanrobot/wholebodycontact", allow_patterns=["datasets/wb30l/*"], local_dir="wbc_artifacts")
# one GR00T fine-tune:
snapshot_download("stevenryanrobot/wholebodycontact", allow_patterns=["groot_n17/n17_sonic_tasks457_force/*"], local_dir="wbc_artifacts")

On-disk locations in the code repo: models/* → data/wbc/mthd2.3/models/, models/mthd2.2_wb30/ → data/wbc/mthd2.2/wb30/models/, models/mthd2.0/ → data/wbc/mthd2.0/models/, models/mthd2.1/ → data/wbc/mthd2.1/, models/grfnet_estimator/ → data/wbc/mthd2.2/grfnet/models/, groot_n17/<run> → ~/wbc_external/ckpt/<run>. Full map: REPRODUCING.md §2 in the code repo.

Sonic and leg-contact datasets (2026-10-28)

Sonic v1.1 collections G3, G4v2, G5, G5v2 and long-contact G6 (325 HDF5 files, 190.469 GB), earlier local Sonic/GRF records, and the existing wb30/wb30l/AMASS/GRF datasets are available. See the data guide for exact version coverage, leg labels, split rules, Pg provenance, and selective downloads. File manifests include sizes and SHA-256 checksums.