Supplementary Information for the Paper "Physical Interaction Effects Reveal the Necessary

Abstract

We conduct a systematic two-year experimental study across 22 contact-rich insertion tasks to determine whether the failure of learning-based robot control on tight-tolerance assembly reflects a fundamental structural limitation. Evaluating five levels of model knowledge across 693,000 real-hardware trials, we show that end-to-end deep reinforcement learning achieves an average success rate of 46.2% even at the highest evaluated level, while a structured Tactile Skill framework encoding robot dynamics, impedance control, and physical interaction effects achieves 100% success on all tasks — learned directly on real hardware in under one hour per task without simulation. Validated across four industrial assembly workflows, the system achieves 150–200% of human daily throughput. The central finding is that physical interaction structure must be explicitly encoded in the learning representation: models matter.