Tactile datasets push robot dexterity past vision-only models

Academic labs and startups are racing to build tactile datasets and training techniques for dexterous robot manipulation, IEEE Spectrum reports. A UC Berkeley team pretrained a vision-language-action model, then trained a specialist tactile submodel on 100 hours of high-quality tactile data covering more than 200 household objects, using separate high-level action and faster low-level tactile experts. Fine-tuned on about 100 teleoperated demonstrations, it averaged 65 percent success across 12 tasks, nearly double the best VLA model. Tsinghua researchers aggregated over 3,000 hours of tactile data across 21 sensor types into a hardware-agnostic model, while Fudan University and spin-out NeoteAI collected more than 30,000 hours. Researchers say how much data is needed remains unclear.