Event

Tsinghua and other teams release HandsOn dexterous manipulation simulation benchmark

2 reports · 2 sources

The team led by Ning Ding of Tsinghua University, together with researchers from Beijing University of Posts and Telecommunications, Peking University, and Nature Will, built HandsOn, a cross-embodiment dexterous manipulation simulation platform and evaluation benchmark. It contains 66 manually verified tasks and 18,000 human simulation teleoperation demonstrations. The platform supports 5 types of robotic arms, 6 types of dexterous hands, and about 200 rigid, articulated, and deformable assets, which can be combined in randomized scenes. The benchmark has two parts: Core-50 is used for learning from demonstration and execution testing, and Transfer-16 evaluates asset and relation transfer, skill composition, and deformable manipulation under the condition of no target-task demonstrations. Without benchmark fine-tuning, GPT-6 used move_eef to control the palm and fingers and, in 660 trials, achieved 5.2% on Core-50 and 1.9% on Transfer-16. Data collection and quantitative evaluation uniformly used the ROKAE—Wuji and PhysX configuration. Real-robot performance currently has only qualitative videos available for observation, and its scope of evidence differs from the quantitative evaluation in simulation.