Simulation-trained robot policy reportedly inserts pegs in 73% of real-world trials

In a paper submitted Oct. 16, Antonia Bronars and colleagues describe assembly policies trained entirely in simulation that use joint positions, velocities and past actions without vision, force or tactile sensing. They report that a peg-insertion policy transferred zero-shot to the real world and succeeded in 73% of trials with varied pegs. The abstract lists gear meshing and nut threading as learned tasks but gives no real-world results for them.