businesslion/physprobe-features

physprobe features: frozen video representations of a pushed cube with hidden mass and friction

Companion data for the code repository https://github.com/msannikov03/physprobe (course project "Do frozen video encoders and a generative video tokenizer see hidden object physics?", Skoltech, 2026; Mikhail Sannikov, Ilya Mikhalchuk). This revision is tagged v1.1, matching the code tag v1.1.

A visually identical 10 cm cube is pushed across a plane in MuJoCo with a fixed impulse. Its mass m (log-uniform in [0.5, 2] kg) and friction coefficient μ (uniform in [0.2, 0.8]) change only how it moves. This dataset holds the frozen-encoder features and baseline blocks of every clip plus the per-clip contact and state signals and metadata, so the linear-probe results can be reproduced on a CPU in about 15 minutes, without regenerating the 42 GB of video or running the encoders on a GPU.

Contents

features/<dataset>/<rep>/<mode>.npz     ids (<U16, metadata order), X (float32, or float16 for sequence/grid modes), info (JSON string)
data/<dataset>/meta.jsonl               one JSON object per clip: id, scenario, m, mu, J, v0, x0_m, seed, split, visibility and stability keys, ok
data/<dataset>/dataset.json             scenario config, camera projection, friction model, git_commit
data/<dataset>/contact_<id>.npy         float32 (64, 3): [fn_N, ft_N, slip_mps] at 64 Hz
data/<dataset>/state_<id>.npy           float32 (64, 6): cube centre [x, y, z, vx, vy, vz] at 64 Hz

Datasets: push (2000 clips; m and μ identifiable), slide (1000 clips, fixed initial speed; only μ identifiable, negative control for mass), poscheck (300 static clips at random positions; checks that pooled features keep position), smoke_push (20 clips; an early smoke set with encoder features only, made with older feature code and used neither by the tests nor by the results).

Feature inventory

push and slide hold 34 files each, poscheck 30 (no noise curve), smoke_push 21 (the encoder rows only, no baseline blocks). Exact definitions, preprocessing and flatten orders: CONTRACT.md in the code repository.

rep	modes	dim	X dtype	oracle
vjepa2_l (V-JEPA 2 ViT-L, facebook/vjepa2-vitl-fpc64-256)	static, clip_mean, clip_mean_L08, clip_mean_L16	1024	float32	no
vjepa2_l	clip_seq, clip_seq_L08, clip_seq_L16	32768 (32 slots × 1024)	float16	no
vjepa2_l_random (same architecture, random weights, fp32)	static, clip_mean	1024	float32	no
vjepa2_l_random	clip_seq	32768	float16	no
dinov2_b (facebook/dinov2-base, per frame), dinov2_b_random	static, last, clip_mean	1536 ([CLS ‖ mean patch])	float32	no
dinov2_b, dinov2_b_random	clip_seq	24576 (16 frames × 1536)	float16	no
cosmos_cv8 (NVIDIA Cosmos-Tokenize1 CV8×8×8 latent)	clip_mean	16	float32	no
cosmos_cv8	clip_seq	144 (9 slots × 16 channels)	float16	no
cosmos_cv8	clip_grid4	2304 (16 × 9 × 4 × 4)	float16	no
pixels	static, last	1024 (grey 32 × 32)	float32	no
pixels	clip_seq	8192 (8 time bins × 32 × 32)	float16	no
track	clip	128 (64 frames × centroid x, y)	float32	no
state	oracle_x (1-px quantised x(t)), oracle_xc (unquantised x(t))	64	float32	yes
state	oracle_fit ([ln v0, a])	2	float32	yes
contact	raw	192 (64 × [fn, ft, slip])	float32	yes
contact	realistic	192	float32	no
contact	noise_000, noise_010, noise_030, noise_100 (push and slide only)	192	float32	no

The raw video clips (uint8, 64×256×256×3 per clip, 42 GB) are not included; scripts/gen_all.sh in the repository regenerates them deterministically in about two minutes on a CPU. Only the two frame cells of the demo notebook need them.

Use

git clone https://github.com/msannikov03/physprobe && cd physprobe && git checkout v1.1
python3 -m venv .venv-probe && . .venv-probe/bin/activate     # Python >= 3.11
pip install -r requirements-probe.txt huggingface_hub==1.20.1
python scripts/download_features.py      # resumable; sets HF_HUB_DISABLE_XET=1 (anonymous downloads of the 6,700 small files hit a 429 rate limit on the xet path otherwise)
scripts/reproduce.sh --from probes             # baselines, probes, tables, gates, figures -> results/; about 15 min on a laptop CPU
python scripts/compare_results.py --rev v1.1   # every probes.csv row against the committed one (|Δ R²| ≤ 1e-6)

allow_patterns keeps this card and its .gitattributes from overwriting the clone's own files. Without clips, the probe step keeps the shipped pixels and track blocks and rebuilds only the state and contact blocks from data/.

Load one feature file:

from physprobe import io
ids, X, info = io.load_features("push", "vjepa2_l", "clip_mean_L16")   # X: (2000, 1024) float32
meta = io.load_meta("push")                                            # list of dicts, same order as ids

Provenance

Generated on 22–23 October 2026 with MuJoCo 3.16.0, torch 2.10.0 and transformers 5.3.0. Model revisions are pinned in physprobe/feats/encoders.py and recorded in each file's info. Hashes inside results/, the feature info and the notebook refer to the private development history that was squashed into the public v1.0 commit d41bcfa; its physprobe/ and tests/ trees are identical. In detail:

• data/ (all four datasets): dataset.json carries git_commit c3dac34, a private commit that predates the simulator code. The data were generated by the simulator code first committed as a124216 (physics identical to the release). These files predate the provenance keys, so they have no sim_code_sha256.
• Encoder features of push, slide and poscheck: info.git_commit fce21d1 (private) and code_hash e86e617ab2fd, which equals the sha1 of encoders.py + extract.py in the public releases v1.0 (d41bcfa) and v1.1.
• Baseline blocks (pixels, track, state, contact) of push, slide and poscheck: info.git_commit is the full SHA of the same private commit fce21d1. The state and contact blocks are rebuilt bit-identically from data/ by the release code.
• smoke_push features are older: 11 files at c3dac34 with code_hash 4fec3256d8ce, and the 3 Cosmos files at f4b5541 with code_hash d0eaf9b73f76, written before CONTRACT v1.3 (no layer or fp32_precision keys).
• Known oddity: slide/vjepa2_l_random/clip_mean and clip_seq carry the release code_hash but lack the fp32_precision key that the release extractor writes. They were probably written by a process that started under older code; this is not confirmed.

Licences

This dataset (metadata, contact and state signals, baseline blocks and feature vectors) is released under CC BY 4.0. The features are outputs of third-party models: V-JEPA 2 (MIT), DINOv2 (Apache-2.0) and Cosmos-Tokenize1. Cosmos-Tokenize1 is Licensed by NVIDIA Corporation under the NVIDIA Open Model License; the cosmos_cv8 features are outputs of that model. No model weights are redistributed here.

Citation

Sannikov M., Mikhalchuk I. physprobe: Do frozen video encoders and a generative video tokenizer see hidden object physics? Course project, Skoltech, 2026. Version 1.1. https://github.com/msannikov03/physprobe (see CITATION.cff there).