T-Rex lifts dexterous hand success to 65% by giving touch its own fast path
量子位 · wechat · 2026-07-26
QbitAI profiles T-Rex, a dexterous-hand system that shows tactile sensing is not optional for fine manipulation.
- On 12 real-world dexterous tasks, T-Rex reaches 65% average success rate, versus 35% for the strongest vision-only baseline.
- Removing tactile input drops performance from 65% to 42%, suggesting touch is carrying a large share of the skill.
- A naive fusion of tactile force signals into a pretrained VLA model hurts performance, with success falling from 17% to 6% in one experiment.
- The system uses a Mixture-of-Transformer-Experts design: a latent expert for vision/language, a large action expert for low-frequency planning, and a lightweight tactile expert for high-frequency refinement.
- Tactile history is compressed with VQ-VAE into discrete codes, and training is done in three stages: large-scale human video pretraining, 100 hours of teleoperation mid-training, and task-specific finetuning with around 100 demos.
- The article argues that dexterous hands are likely to become a standalone embodied-AI segment, with tactile data becoming a key moat.
More from Embodied
- Rumor says Flux 3 unifies image, video, audio, and action prediction — theteknosaur · 2026-07-26
- Chinese robotics team trains humanoids with kicks, pushes, and knockdowns — 2C_ornot2C · 2026-07-26
- D-Propulse previews an air-breathing rotating detonation engine with aerospike nozzle — Paimaamu · 2026-07-26
- A reMarkable Paper Pro demo turns an e-ink tablet into an AI notebook — Olivier__OG · 2026-07-26
- Chinese robot dog switches between wheels and legs autonomously — RoboBalaji · 2026-07-26
- Ant Group releases 2,000 hours of egocentric manipulation data for embodied learning — burny_tech · 2026-07-26