Camera-free robot policy shows a reported 73% success rate in real-world peg-insertion trials

Researchers say their simulation-trained robot policies use joint positions, velocities and past actions instead of visual, force or tactile sensors. They report 73% success on real-world peg insertion across parts of different shapes and sizes; gear meshing and nut threading were demonstrated in simulation. In simulated comparisons, blind policies generalized better than matched visual policies to unseen parts, though they could be slower. The researchers also released mjfit, an open-source generator of MuJoCo assembly models.