Robots already use their own joint sensors to walk and track whole-body motion, but those contacts are large and forceful. The authors of this paper ask whether the same proprioceptive sensing is enough for delicate, contact-rich assembly—finding a millimeter-scale opening, meshing gear teeth, or catching a thread. They trained "blind" policies with reinforcement learning entirely in simulation, giving the policies only a short history of joint positions, velocities, and actions. The pose of the part and fixture were unknown, and grasps were randomized, so the robot had to infer everything through the feel of contact.
The resulting policies learned peg insertion, gear meshing, and nut threading. The peg-insertion policy transferred zero-shot to the real world and succeeded in 73% of trials across different shapes and sizes. In controlled simulation comparisons, the blind policies generalized far better to unseen part shapes and sizes than matched visual policies, though they could be slower. This suggests that proprioceptive contact, rather than vision, may be a surprisingly strong signal for generalization in precise manipulation.
The authors also released an open-source procedural generator that builds thousands of assembly tasks with exact convex pieces for the MuJoCo simulator. They report it is more accurate and faster than convex decomposition or MuJoCo's signed-distance collision. The work is a single study, so there are no conflicting sources to compare, but the results point toward simpler, sensor-light robot assembly pipelines that may transfer robustly from simulation to reality.