HandsOn benchmark: best dexterous manipulation policy manages only 29.1% success on Core-50
jiqizhixin · x · 2026-11-05
Researchers from Tsinghua, BUPT, PKU and NatureWill release HandsOn, a simulation benchmark for bimanual dexterous manipulation pairing 18,000 curated human demonstrations with 66 human-verified tasks. Core-50 tests demonstration-backed execution while Transfer-16 tests generalization to new assets, skill combos and deformable objects. Best multitask policy hits just 29.1% task-average success on Core-50 (GPT-6 prompted VLM: 5.2%), with failures concentrated in handover timing, bimanual desynchronization and goal maintenance.
More from Embodied
- What if Boston Dynamics just made boring robot arms and mobile manipulators? — lukas_m_ziegler · 2026-11-05
- NVIDIA's TAMPAgent uses factor graphs to let robots zero-shot mobile manipulation in unseen homes — danfei_xu · 2026-11-05
- YC F26 Startup Asim Launches Benchmarks to Find Robot Failures Before Deployment — ycombinator · 2026-11-05
- MachineAge launches marketplace paying skilled trades to train robots — thetripathi58 · 2026-11-05
- OpenCV Live Demos a Robot 'Visual Cortex' with RealSense, Jetson, ROS 2 and VLMs — chrismatthieu · 2026-11-05
- FoveaVLA lets the robot policy choose where to look: 70.8% vs 52.1% for wrist cams — stepjamUK · 2026-11-05