T-Rex lifts dexterous robot success to 65%

NVIDIA and UC Berkeley’s open T-Rex policy reached a 65% macro-average success rate across 12 tactile-reactive manipulation tasks, 30 percentage points above its strongest baseline. The project page and the paper release a 100-hour tactile dataset and a variable-rate model that runs touch corrections faster than visual planning; the benchmark uses 16 randomized trials per task and a 58-DoF bimanual robot. Robot teams should treat tactile data and high-frequency control as a joint systems investment, extending the archive’s evidence that physical-AI pilots need measured reliability.