Megadose The AI curve, compressed.

Correcting Persistent Failures in Point-Cloud Grasping Policies using Human Demonstrations in Virtual Reality

· HF Daily Papers ·
Just 58 human demonstrations raised simulated grasping success on persistent failure cases from 24% to 45%.

The system automatically identifies grasps a reinforcement-learning policy repeatedly fails and sends them to a virtual-reality application for human correction. A person supplies two grasp points; the robot chooses its own approach. The demonstrations also recovered performance on similar cases that were never demonstrated, without reducing success on unseen objects. All reported results were in simulation.

HF Daily Papers' note

score 4

Categories: Research