mcobzarencoapache-2.0

Community Curated v0

A filtered, quality-annotated union of 981 crowd-sourced robot manipulation datasets containing 52,507 episodes and 24.8M frames (~229 hours) of real-world teleoperation on 6-DoF hobby-class arms, all annotated with vision-language judge verdicts, task instructions, subgoals, and per-frame progress labels.

Downloads39
Episodes52,507
Hours229

Why This Matters for Physical AI

This large-scale, heterogeneous dataset with dense vision-language annotations provides a valuable resource for training and evaluating imitation learning models on diverse real-world manipulation tasks across multiple robot platforms and conditions.

Technical Profile

Modalities
rgbproprioceptionlanguage
Robot Embodiments
SO-100SO-101
Action Space
joint_positions
Environment
lab
Task Types
manipulationgraspingpick_and_place
Episodes
52,507
Total Hours
229
Data Format
LeRobot
Annotation Types
language_instructionsreward_labelsaction_labelssegmentationbounding_boxes
License
apache-2.0
Part of the LeRobot family

Access

Need custom rgb data?

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

Request a Sample Pack

Related Datasets