The study investigates if joint sensing is sufficient for precise assembly tasks, such as inserting pegs into holes, meshing gears, and threading nuts, which require finding small openings and making light contacts.
The authors propose training blind assembly policies using only joint position, velocity, and action history, without any visual, force, torque, or tactile inputs. These policies must learn to assemble parts through contact, relying solely on proprioception.
The policies were trained via reinforcement learning in simulation with randomized parts, grasps, and contacts. They demonstrated successful peg insertion in 73% of real-world trials. In simulation, blind policies showed better generalization to unseen part shapes and sizes compared to visual policies, though they may be slower.