Spa-Bench2026apache-2.0
Spa-Bench Teleoperated Training Dataset — 1,200 Episodes
A teleoperated demonstration dataset of 1,200 episodes with five synchronized camera streams used for compositional generalization experiments across six task families involving manipulation and object interactions.
Downloads223
Episodes1200
Hours5.6735
Why This Matters for Physical AI
This dataset enables training and evaluation of vision-language-action models for compositional generalization in robotic manipulation, where policies must generalize to unseen concept-argument combinations while maintaining semantic understanding of task instructions.
Technical Profile
- Modalities
- rgbproprioception
- Robot Embodiments
- SO-101
- Action Space
- 6-DoF: shoulder_pan.pos, shoulder_lift.pos, elbow_flex.pos, wrist_flex.pos, wrist_roll.pos, gripper.pos
- Environment
- home
- Task Types
- manipulationgraspingpick_and_placecountingrelational_placement
- Episodes
- 1200
- Total Hours
- 5.6735
- Data Format
- LeRobot
- Annotation Types
- language_instructionsaction_labels
- License
- apache-2.0
Access
Need custom rgb data?
Claru builds purpose-built datasets for home applications with dense human annotations and quality assurance.
Request a Sample Pack