PRISM: Precision and contact-rich Real-world Industrial Skill Dataset with Multimodal Sensing
PRISM is a large-scale multimodal dataset for contact-rich real-world industrial manipulation. It is designed to support robot learning in industrial scenarios where precise contact, force/torque regulation, tactile feedback, and multimodal perception are important.
Most existing robot learning datasets focus on short-horizon and low-contact manipulation tasks, such as pick-and-place. These datasets are often insufficient for industrial assembly and other contact-rich operations, where robots must handle tight tolerances, sustained contact, friction, insertion, alignment, and force-sensitive interactions.
PRISM addresses this gap by collecting diverse industrial manipulation demonstrations with synchronized multimodal sensing. The dataset includes multi-view RGB-D observations, force/torque measurements, tactile signals, and robot-state information, providing a realistic benchmark for learning contact-rich manipulation policies.
- 25+ industrial manipulation tasks
- 5,000+ robot trajectories
- 5,000+ paired human demonstrations
- 45+ hours of demonstrations
- ~27M images across visual and visuotactile streams
- Multi-view RGB-D sensing
- 6DoF force/torque sensing
- Tactile and proprioceptive observations
- Multiple robot platforms and teleoperation interfaces
BibTeX will be released soon.