CLARU
Rongxuan-ZhouMIT

OGBench Scene expert episodes

3,000 expert episodes (606,000 frames) of block manipulation with buttons, drawer, and window in the OGBench Scene environment, collected via scripted expert through the stable_worldmodel pipeline.

Downloads11
Episodes3000

Why This Matters for Physical AI

This dataset provides high-quality expert demonstrations for training world models in manipulation tasks, supporting research in predictive models for robotic control.

Technical Profile

Modalities
rgbproprioception
Robot Embodiments
robotic_arm
Action Space
joint_positions
Environment
simulation
Task Types
manipulationobject_interaction
Episodes
3000
Data Format
HDF5
Annotation Types
action_labelsprivileged_state
License
MIT
Part of the OGBench family

Access

View on HuggingFace

Need custom rgb data?

Claru builds purpose-built datasets for simulation applications with dense human annotations and quality assurance.

Request a Sample Pack

Related Datasets

OGBench mirror: cube-quadruple-play-100m-v0

Unofficial mirror of the cube-quadruple-play-100m-v0 dataset from OGBench, a benchmark for offline goal-conditioned reinforcement learning. The dataset contains 100 million steps of simulated robotic manipulation data.

501 downloadsSep 2026MIT

OGBench: cube-triple-play-100m-v0

Unofficial mirror of the cube-triple-play-100m-v0 dataset from OGBench, a benchmark for offline goal-conditioned reinforcement learning. Contains 100 million steps of simulated robot manipulation data for cube manipulation tasks.

355 downloadsSep 2026mit

OGBench mirror: cube-double-play-v0

Unofficial mirror of the cube-double-play-v0 dataset from OGBench, a benchmark for offline goal-conditioned reinforcement learning. The dataset contains trajectories of a robotic system performing cube manipulation tasks in simulation.

73 downloadsSep 2026mit

OGBench mirror: antmaze-large-navigate-v0

Unofficial mirror of the antmaze-large-navigate-v0 dataset from OGBench, a benchmark for offline goal-conditioned reinforcement learning. The dataset consists of simulated ant navigation trajectories in a large maze environment.

proprioception
71 downloadsSep 2026mit

OGBench humanoidmaze-medium-navigate-v0

Unofficial mirror of the humanoidmaze-medium-navigate-v0 dataset from OGBench, a benchmark for offline goal-conditioned reinforcement learning. Contains trajectory data for training offline RL agents on humanoid navigation tasks in maze environments.

proprioception
70 downloadsSep 2026MIT

OGBench mirror: puzzle-4x4-play-v0

Unofficial mirror of the puzzle-4x4-play-v0 dataset from OGBench, a benchmark for offline goal-conditioned reinforcement learning. The dataset contains trajectories for a 4x4 puzzle manipulation task in simulation.

69 downloadsSep 2026mit