OSCAR_human

OSCAR companion data — egocentric-human split.

This repository hosts the egocentric-human half of the training corpus for OSCAR: Omni-Embodiment Action-Conditioned World Model for Robotics, a precise action-conditioned video world model for cross-embodiment robot policy evaluation. It holds egocentric human-manipulation episodes (MANO hand) that we curated, filtered, deduplicated, and re-rendered into OSCAR's unified conditioning format — the same as the robot split: a rendered MANO-skeleton overlay aligned to the original RGB, an OSCAR-generated caption, and numeric per-frame metadata. Human video adds scene and motion diversity that robot teleoperation alone does not provide. Two upstream sources (EgoDex, VITRA/Epic-Kitchens); all clips are 70 frames. The robot split lives in zywu2115/OSCAR_robot.

• Paper: OSCAR: Omni-Embodiment Action-Conditioned World Model for Robotics
• Authors: Zhuoyuan Wu (Peking University), Jun Gao (University of Michigan; NVIDIA).
• Project page: https://wuzy2115.github.io/oscar-project-page/

Per-subset license

Subset	Upstream license	Notes
egodex	CC-BY-NC-ND-4.0	⚠️ NC
vitra_epic	CC-BY-NC-4.0	⚠️ NC

What's inside

Total episodes: 85827

• egodex: 78273 episodes — EgoDex: Egocentric Dexterous Manipulation
• vitra_epic: 7554 episodes — VITRA/Epic-Kitchens Egocentric Hand Manipulation

Why no rgb.mp4?

Upstream RGB videos are subject to each upstream dataset's license (see the per-subset table above). To respect each upstream's distribution terms, we redistribute only our derived artifacts: the rendered skeleton overlay, the captions we generated, and the numeric episode metadata.

To recover RGB frames, see Mapping back to upstream.

Dataset structure

<root>/
├── egodex/
│   └── ...episode dirs, each containing:
│       ├── caption.pickle            # OSCAR-generated
│       ├── episode_meta.npz          # numeric meta (see schema below)
│       └── skeleton_scenario.mp4            # OSCAR-rendered skeleton overlay
├── vitra_epic/
│   └── ...episode dirs, each containing:
│       ├── caption.pickle            # OSCAR-generated
│       ├── episode_meta.npz          # numeric meta (see schema below)
│       └── skeleton_scenario.mp4            # OSCAR-rendered skeleton overlay

Per-file schema

caption.pickle

Pickle dict with keys: text (str), model (str, our caption model id), version (str).

episode_meta.npz

Field	Shape	Dtype	Origin	Semantics
joint_positions_3d	(T, J, 3)	float64	OSCAR-derived from upstream	SMPL-X / MANO FK applied to upstream pose params
camera_intrinsic	(3, 3)	float64	upstream passthrough	pinhole K
camera_extrinsic	(T, 4, 4)	float64	upstream passthrough	per-frame head-mounted camera pose
camera_is_per_frame	()	bool	OSCAR bookkeeping	True for human (always)
visible	(T,)	bool	OSCAR-generated	per-frame visibility flag from our filter

skeleton_scenario.mp4

H.264 (libx264), frame-count matches the (deleted) rgb.mp4. Contains the OSCAR-rendered skeleton overlay. Per-link / per-finger color conventions follow OSCAR's internal skeleton_def.py.

Filters applied

human filters: min_frames=70, min_visible_ratio=0.3, min_hand_motion=0.005

Mapping back to upstream

Episodes here preserve the upstream's relative paths and episode identifiers (with one caveat per subset).

1. Identify the subset (top-level directory).
2. Match the relative path against the upstream release.
   • See each subset README for the exact naming convention and the upstream download link.
3. Download the upstream raw to fetch RGB and align with our skeleton render.

We plan to release a helper script that automates upstream-raw download + RGB alignment in a follow-up release. Until then, please follow the per-subset instructions linked below.

License

Each subset inherits its upstream license. See the subset README for canonical license text.