Event

Blind assembly policy simulation training achieves 73% real-world peg insertion success rate

1 reports · 1 sources

Antonia Bronars and other researchers propose a blind assembly policy that relies solely on proprioception, using no vision, force, torque, or tactile sensing, and observing only a short history of joint positions, velocities, and its own actions; in simulation, reinforcement learning is used to complete peg insertion, gear meshing, and nut tightening tasks. The peg insertion policy transfers zero-shot to the real world, successfully inserting pegs of different shapes and sizes in 73% of trials. Policy training is conducted entirely in simulation, with randomization over parts, grasps, and contacts, and the policy is unaware of the poses of both the fixture and the parts, with a relatively high degree of grasp randomization. Controlled experiments show that the blind policy generalizes better to unseen part shapes and sizes, but may be slower. The researchers also released an open-source procedural assembly generator that builds tasks in MuJoCo based on exact convex blocks.