Learning to Unscrew the Cap Naturally: Skill Decomposition and Policy Distillation for Robot Twisting

Abstract

Dexterous twisting is a fundamental yet challenging capability for anthropomorphic manipulation due to its high-dimensional action space, strong spatio-temporal coupling, and complex contact dynamics. While reinforcement learning (RL) often struggles with convergence under such conditions, learning from demonstration (LfD) requires extensive data collection and suffers from human-to-robot retargeting errors. This paper proposes SDPD, a task-oriented learning method that integrates skill decomposition (SD-RL) and policy distillation (PD). The SD-RL module decomposes the twisting task into three sequential sub-skills: Grasp, Rotate, and Release, and learns physically consistent policies under tactile guidance. Subsequently, an offline policy distillation procedure based on TD3+BC consolidates stage-wise expert policies into full-process policies with reduced dependence on hard phase switching. Simulation results show that SDPD improves manipulation efficiency and long-horizon twisting performance while maintaining coordinated contact-rich motions compared with representative baselines. Across the tested cap-size range, the proposed method achieves high-performance execution over 89.1% of the range, improving the coverage by 51.4 percentage points over the baseline. Sim-to-real deployment on the Shadow Hand platform further demonstrates stable real-world performance: the system can continuously twist the bottle cap for more than 25 full turns over a 20 -minute period while maintaining a motion success rate above 85%.