Silicon23/mujoco-kinematics-probing

MuJoCo Kinematics — V-JEPA 2 Probing Dataset

Synthetic physics videos with exact per-frame ground-truth labels, generated to probe whether video world models (V-JEPA 2 in particular) encode kinematic and physical properties in their latent representations.

Research project: Do Video World Models Encode Kinematics? (CSE 493S).

Contents

• 10,000 clips across 5 scenarios (2,000 clips each)
• Every clip: 4 seconds × 16 fps × 256×256 (64 frames) — matched to V-JEPA 2's input format
• Per-clip: clip.mp4, labels.npz (per-frame ground truth), meta.json (config + sampled parameters)
• Plain high-contrast background, fixed camera auto-framed to the trajectory, no shadows — the only decodable signal is the object(s) and their motion

Scenarios

name	objects	description
constant_velocity	1	straight-line motion, zero gravity
free_fall	1	motion under varied gravity (magnitude + direction sampled per clip)
constant_acceleration	1	motion under arbitrary constant acceleration
wall_collision	1 + fixed wall	object bounces off a fixed wall
ground_bounce	1 + floor	object falls and bounces on the floor

Files

manifest.jsonl              one line per clip with key metadata
dataset_config.json         run configuration (fps / resolution / duration / seed)
previews/<scenario>.png     contact-sheet preview (64 frames tiled) per scenario
constant_velocity.tar       2000 clips    (~70 MB)
free_fall.tar               2000 clips    (~70 MB)
constant_acceleration.tar   2000 clips    (~70 MB)
wall_collision.tar          2000 clips    (~70 MB)
ground_bounce.tar           2000 clips    (~70 MB)

Each tar extracts to <scenario>/clip_NNNNNN/{clip.mp4, labels.npz, meta.json}.

Download and extract

from huggingface_hub import snapshot_download
import tarfile, pathlib

local = snapshot_download("IOAI-ISC/mujoco-kinematics-probing", repo_type="dataset")
for tar in pathlib.Path(local).glob("*.tar"):
    with tarfile.open(tar) as tf:
        tf.extractall(local)

Label schema (labels.npz)

N = objects in clip, T = 64 frames. Per-object arrays are indexed (N, T, …).

key	shape	meaning
time	(T,)	simulation time per frame
pos	(N,T,3)	object centre, world position
quat	(N,T,4)	orientation (wxyz)
lin_vel	(N,T,3)	linear velocity, world frame
ang_vel	(N,T,3)	angular velocity, body frame
pix	(N,T,2)	object centre projected to pixel (u,v)
in_frame	(N,T)	object centre inside frame bounds
n_contacts	(T,)	contacts at the frame instant
contact	(T,)	any contact since the previous frame
mass	(N,)	object masses (kg)
size	(N,)	object sizes (m)
momentum	(T,3)	total linear momentum
kinetic_energy	(T,)	total kinetic energy (translational + rotational)
potential_energy	(T,)	−Σᵢ mᵢ (g·rᵢ), reference z = 0

meta.json adds: scenario, seed, render config, objects (per-object physical properties + observability flags), sampled params (including the gravity vector, gravity_magnitude, gravity_class, gravity_reversed, and scenario-specific values like restitution and mass_ratio), derived (measured restitution), and full scene description.

Gravity distribution

Scenarios that use gravity (free_fall, ground_bounce, constant_acceleration) sample |g| from a mixture matching the probing categories:

• 70% Earth range — uniform on [5, 15] m/s²
• 15% OOD-low — log-uniform on [1, 5) m/s²
• 15% OOD-high — log-uniform on (15, 50] m/s²

free_fall flips gravity direction 50/50 (normal vs reversed) for the gravity-direction classification probe. ground_bounce keeps gravity downward (reversed gravity is physically incompatible with floor contact).

Observability caveats

A probe recovers a property from pixels. Some labels have no visual trace given the object — the probe cannot recover them by physics, not by model failure. Use the metadata flags to filter probes:

• Angular velocity / orientation / moment of inertia — invisible for a rotationally_symmetric object (uniform sphere). Every object in this dataset is a uniform sphere, so these probes are not meaningful as-is.
• Absolute mass / momentum / kinetic energy — not recoverable for a single free body (all masses fall identically under a uniform gravitational field). Recoverable only in collision scenarios — gate on has_collision.
• Friction — only acts during sliding/rolling contact.
• Restitution — only observable if a bounce occurs in-clip; use derived.restitution_measured as the exact label (the params.restitution requested value is a sampling knob).
• Depth / absolute size — monocular scale ambiguity. The generator keeps motion in the image plane (depth y = 0) to sidestep this.

Citation

Generated using MuJoCo 3.8.1. Generation code: see the project repository.