Pushing Dexterity and Safety for In-Hand Object Pose Reaching That Works Remarkably Well
A learned safety filter cut object drops by 75%, while the robot hand still reached 91.5% of target poses.
The researchers trained a LEAP hand in simulation without human data, then transferred its policies to hardware without further training. Across 13.3 hours of testing, the state-based policy reached 95.2% of 4,800 unseen goals across four objects within tolerances of 2 cm and 0.4 radians before adding the filter. The filter’s adjustable conservativeness lets operators trade some reaching performance for fewer drops. The policies also worked with a single wrist-mounted depth camera and demonstrated tasks including screwing in a lightbulb.
ArXiv · AI/CL/LG's note
The researchers trained a LEAP hand in simulation without human data, then transferred its policies to hardware without further training. Across 13.3 hours of testing, the state-based policy reached 95.2% of 4,800 unseen goals across four objects within tolerances of 2 cm and 0.4 radians before adding the filter. The filter’s adjustable conservativeness lets operators trade some reaching performance for fewer drops. The policies also worked with a single wrist-mounted depth camera and demonstrated tasks including screwing in a lightbulb.
ArXiv · AI/CL/LG's note
score 5