Jim Fan open-sources T-Rex: giving robots a first-class sense of touch
DrJimFan · x · 2026-08-22
NVIDIA's Jim Fan shared his team's newly open-sourced methodology "T-Rex", arguing touch is the most criminally under-explored modality in robotics — cameras can't see a magnetic piece snapping into place or a USB negotiating into its port, like doing sleight of hand in oven mitts.
T-Rex treats tactile as a first-class citizen: a mixture-of-transformers runs two clocks asynchronously, where a slow visuomotor expert plans motion and a fast tactile expert refines it with high-frequency real-time corrections.
Related event: NVIDIA and UC Berkeley Open-Source T-Rex Tactile Robot Learning(2 posts)→
More from Embodied
- Humanoid Robot Crashes at Beijing 'Robot Olympics' Sprint — steveplunkett · 2026-08-22
- Deep Dive: How World Models Unlock Physical AI and Robotics — risingodegua · 2026-08-22
- RoboRobots Cuts Robot Deployment Costs to $10/hr, Seeks Fleet Partners — ai · 2026-08-22
- Humanoid robots walk in air and form bands at World Humanoid Robot Games rehearsal — CyberRobooo · 2026-08-22
- Robotics gatekeeping: No motor/driver build, no entry — wavefnx · 2026-08-22
- DIY Coil Winder: Building Custom Actuators from Scratch — IanPritchard · 2026-08-22