NVIDIA T-Rex Tactile Robotics Open Source

NVIDIA and UC Berkeley open-sourced T-Rex, a tactile robotics methodology that treats touch as a first-class modality. The architecture uses a mixture-of-transformer with an asynchronous slow visuomotor expert and a fast tactile expert running at 4 touch ticks per vision tick. They released a 50-hour tactile dataset (~5,500 episodes) on HuggingFace, collected on a state-of-the-art tactile hand with 22 degrees of freedom. The training recipe extends EgoScale, using human egocentric video pretraining and tactile robot play mid-training to bridge contact-free pretraining to contact-rich manipulation.
Key insights
T-Rex advances robotic manipulation by integrating high-frequency tactile feedback into the model architecture
NVIDIA and Berkeley's T-Rex framework makes tactile sensing a first-class input, with a fast tactile expert running at 4× the vision rate to refine motions in real time. The release includes the largest tactile dataset ever (50 hours, ~5,500 episodes) on a 22-DOF hand. The training approach uses human egocentric videos for pretraining and tactile play for mid-training, enabling better contact-rich tasks. This addresses the gap where vision fails at the moment of contact.