李飞飞第一篇「触觉」论文
Merged summary
TL;DR - T-Rex is a tactile-visual-language-action system for dexterous robotic hands that separates low-frequency planning from high-frequency tactile control. It achieves 65% average success across 12 real-world tasks, versus 35% for the strongest vision-only baseline.
- A Mixture-of-Transformer-Experts architecture uses separate latent, action, and tactile experts, with cascaded flow matching injecting touch feedback during action refinement.
- A VQ-VAE encodes 16-frame force histories into 64 discrete codes, reducing sensor noise while representing temporal contact patterns.
- Removing tactile input lowers average success from 65% to 42%; direct concatenation of force signals into a pretrained VLA model instead reduced success from 17% to 6%.
- Three-stage training combines 22,889 hours of human video, 100 hours of robot teleoperation, and roughly 100 task-specific demonstrations.
Sources (1)
李飞飞第一篇「触觉」论文
TL;DR - T-Rex is a tactile-visual-language-action system for dexterous robotic hands that separates low-frequency planning from high-frequency tactile control. It achieves 65% average success across 12 real-world tasks, versus 35% for the strongest vision-only baseline.
- A Mixture-of-Transformer-Experts architecture uses separate latent, action, and tactile experts, with cascaded flow matching injecting touch feedback during action refinement.
- A VQ-VAE encodes 16-frame force histories into 64 discrete codes, reducing sensor noise while representing temporal contact patterns.
- Removing tactile input lowers average success from 65% to 42%; direct concatenation of force signals into a pretrained VLA model instead reduced success from 17% to 6%.
- Three-stage training combines 22,889 hours of human video, 100 hours of robot teleoperation, and roughly 100 task-specific demonstrations.