Sim-to-Real Reinforcement Learning for Vision-Based Dexterous Manipulation on Humanoids
Average success was 62.3% for grasp-and-reach, 80% for box lift, and 52.5% for bimanual handover.
The policies showed qualitative zero-shot transfer to unseen objects and robustness to force disturbances.
Hours between intervention, repeated-task success over hundreds of trials, cycle time, hand durability, service cost, and paid deployment.
Average success was 62.3% for grasp-and-reach, 80% for box lift, and 52.5% for bimanual handover.
p. 8, System CapabilitiesThe policies showed qualitative zero-shot transfer to unseen objects and robustness to force disturbances.
p. 8, System CapabilitiesThe authors say the system is far from general-purpose human manipulation.
Dynamics transfer uses basic domain randomization and may explain the weak handover result.
Reliable, sufficiently dexterous hardware remains a binding constraint.
Sim-to-real progress lowers the software barrier, but 52.5-80% task success is not commercial reliability.
Thematic