large-scale tactile demonstration dataset and robot-free data collectionIntermediate
TacClaw-Infinity: A Large-Scale Vision-Tactile-Language-Action Dataset for Dexterous Manipulation
Daimon Robotics and partner universities built 200 handheld grippers with fingertip touch sensors and used them to record more than 10,000 hours and 1 million demonstrations of people handling objects, without a robot in the loop. In this preprint, adding the touch signal lifted the average success of robot policies trained only on that data from about 50% to about 74% on three real two-arm tasks.