DexAgent aims to learn two-handed robot manipulation from a single human video

DexAgent’s team introduced a framework for learning dexterous, two-handed robot manipulation from a single human video. Across 11 long-horizon tasks involving rigid, articulated and deformable objects, the team reports a 63.6% policy rollout success rate versus 18.2% for the strongest baseline. It also reports average inference time of 2.1 hours with its tool library, compared with 3.3 hours for the baseline. The library contains 103 skills and 188 verifiers.