Tsinghua Opens HandsOn Benchmark; Best Policy Hits 29.1% on Core-50

The best trained multi-task policy on a new Tsinghua-led dexterous-manipulation benchmark succeeds on 29.1% of its 50 core tasks. HandsOn, from researchers at Tsinghua University, Beijing University of Posts and Telecommunications, NatureWill and Peking University, pairs 18,000 curated human demonstrations with 66 human-verified tasks across five robotic arms, six dexterous hands and roughly 200 rigid, articulated and deformable assets.

The top score belongs to OpenWAM, a world-action model. The paper's rank list runs OpenWAM at 29.1%, FastWAM at 26.4% and π₀.₅ at 24.9%. Only 11 of the 50 Core-50 tasks cross 50% success for the best trained policy.

Prompted frontier models fare much worse. "GPT-6 using move_eef for palm-and-finger control achieves 5.2% on Core-50 and 1.9% on Transfer-16 over 660 trials," the authors write — 29 successes in total, across 10 Core-50 tasks and 2 Transfer-16 tasks.

The failure taxonomy is specific. The paper reports "failures in contact establishment, action coordination, and maintaining achieved goals, with additional sequencing errors in composed tasks." Button and trigger presses fire before fingers are in position: in Photograph Logo, 774 of 778 episodes press the shutter before aiming. In Set Dual Knobs, both hands reach the target in 38 of 800 episodes but hold it at once in only 5.

Skill group matters more than any headline average. Nonprehensile pushing and rolling hits 51.3%. Articulated-part manipulation reaches 42.9%. Grasp–transport–release lands at 34.9%. Tool-on-object sits at 2.9%, precision fit at 2.2%. Deformable manipulation on Transfer-16 scored 0% across checkpoints.

Code is at github.com/TsinghuaC3I/HandsOn, and the authors ran everything on NVIDIA Isaac Sim 5.1. Sim-to-real evidence is limited: the abstract offers only "qualitative real-robot rollouts" illustrating transfer.