Blind Precise Assembly: Proprioceptive Sim-to-Real Peg-in-Hole for Diverse Parts

TL;DR

The paper presents a method for training robots to perform precise assembly tasks without vision or external sensors, relying solely on joint position, velocity, and action history. Policies trained in simulation using procedurally generated parts generalize well to real-world blind assembly of diverse pegs, gears, and nuts, achieving 73% success on unseen shapes and sizes. The approach uses reinforcement learning with proprioceptive feedback and a new open-source generator for accurate, fast collision modeling in MuJoCo.

Why It Matters

This enables robust, low-cost robotic assembly in environments where vision or force sensors are impractical or unavailable. The ability to generalize across diverse parts without retraining reduces deployment barriers for industrial automation. It demonstrates that proprioception alone can support high-precision manipulation, opening new paths for sensor-agnostic robot control.

Key Findings

• Policies trained with only proprioceptive feedback can successfully perform complex assembly tasks like peg insertion, gear meshing, and nut threading in simulation.
• Zero-shot transfer to real-world hardware achieves 73% success rate across varied peg geometries and sizes, outperforming visual policies in generalization.
• A procedurally generated assembly dataset with accurate convex collision modeling improves simulation efficiency and fidelity over traditional convex decomposition methods.