HandsOn benchmark targets cross-embodiment dexterous manipulation
A new arXiv paper introduces HandsOn, a composable simulation platform and benchmark for cross-embodiment dexterous manipulation, spanning five robotic arms, six dexterous hands and roughly 200 rigid, articulated and deformable assets in randomized scenes. It pairs 18,000 curated human demonstrations with 66 human-verified tasks, split into Core-50 for demonstration-based learning and execution and Transfer-16 for asset and relation transfer, skill composition and deformable manipulation without target-task demonstrations. The authors report that the best trained multi-task policy reaches 29.1% success on Core-50, while an un-finetuned GPT-6 using move_eef for palm-and-finger control scores 5.2% on Core-50 and 1.9% on Transfer-16 over 660 trials; stage diagnostics point to failures in contact establishment, action coordination and maintaining achieved goals, plus sequencing errors in composed tasks, and sim-to-real transfer is illustrated only by qualitative real-robot rollouts.
The authors say HandsOn offers a shared basis for investigating manipulation generalization and the limits of reliable physical interaction.
Computer Science > Robotics
arXiv:2610.26765 (cs)
[Submitted on 19 Oct 2026]
Title:HandsOn: A Large-Scale Simulation Benchmark for Generalizable Dexterous Manipulation
Authors:Tianwei Luo, Yiming He, Hongjin Wu, Ting Pan, Debin Ren, Haorui Guo, Junlin Yang, Yuxin Zuo, Zhehao Kou, Kaisen Yang, Zhuoqun Lai, Yi Ding, Jiale Yu, Yuchen Zhang, Zijie Qiu, Haozhan Li, Ning Ding
View a PDF of the paper titled HandsOn: A Large-Scale Simulation Benchmark for Generalizable Dexterous Manipulation, by Tianwei Luo and 16 other authors