HandsOn benchmark measures cross-embodiment dexterous manipulation
Original title: HandsOn: A Large-Scale Simulation Benchmark for Generalizable Dexterous Manipulation
The best trained multi-task policy reached 29.1% Core-50 success; untuned GPT-6 scored 5.2%.
HandsOn is a simulation platform and benchmark for dexterous manipulation across embodiments, objects, and tasks. It covers five arms, six hands, about 200 assets, 18,000 human demonstrations, and 66 verified tasks split into Core-50 and Transfer-16. Reported success stays low for trained policies and prompted models, while qualitative real-robot rollouts illustrate sim-to-real transfer.
- Transfer-16 tests asset transfer, skill composition, and deformable manipulation.
- GPT-6 scored 1.9% on Transfer-16 over 660 trials without benchmark finetuning.
- GPT-6 used move_eef for palm-and-finger control.
- Diagnostics cite contact, coordination, goal maintenance, and sequencing failures.
From the original article
1 min read at huggingface.co
General-purpose manipulation requires reliable behavior across embodiments, objects, and task demands. We introduce HandsOn, a composable simulation platform and benchmark for cross-embodiment dexterous manipulation. It supports five robotic arms, six dexterous hands, and approximately 200 rigid, articulated, and deformable assets in randomized scenes.
By Tianwei Luo, Yiming He, Hongjin Wu · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.