Berkeley & Nvidia's T-Rex Framework Lets Robots Feel Objects — No Touch Sensors Required
Robots no longer need to rely on vision alone — they can now feel.
Deep Dive
Researchers from UC Berkeley, Nvidia, and Stanford built a 100-hour tactile-synchronized teleoperation dataset covering 200+ everyday objects and 22 motor primitives. They used ManusMeta gloves to capture finger motion, retargeted onto SharpaRobotics Wave dexterous hands for bimanual teleoperation. Code available at tactile-rex.github.io.
Key Points
- 100-hour tactile-synchronized dataset covers over 200 everyday objects and 22 motor primitives.
- Uses ManusMeta gloves and SharpaRobotics Wave dexterous hands for precise finger motion capture and retargeting.
- Code and dataset publicly available at tactile-rex.github.io for community use.
Why It Matters
T-Rex enables robots to physically feel objects, transforming human-robot interaction and precision tasks.