Reliability and safety
What Separate Goal and Drop Metrics Showed in LEAP-Hand Tests
A real-robot preprint reports how a residual safety filter changed pose-goal attainment and object drops across over 4,800 unseen goals.
A dexterous hand may reach a requested object pose yet still lose the object, so the LEAP-hand preprint treated pose-goal attainment and drops as distinct outcomes. The authors evaluated a state-based policy on a real robot for 13.3 hours, covering over 4,800 unseen goals across 4 objects and diverse wrist orientations. A goal counted as reached when the object came within 2 cm and 0.4 rad of its target pose.
In the authors’ real-robot results, the state-based policy reached 95.2% of pose goals within 2 cm and 0.4 rad; after the residual safety filter was added, drops fell by 75% while goal attainment remained 91.5%.
The available abstract does not give absolute drop counts or rates, results for each object, or measurements for different settings of the adjustable filter. These are therefore author-reported preprint results for this LEAP-hand setup, not independent confirmation or a general safety prescription for dexterous robots.