Structural variability and hidden contact interactions in robotic disassembly impede conventional task-specific control and learning paradigms. This paper formulates twist-pull (TP) as a fundamental execution skill primitive, a generic interaction pattern underpinning diverse disassembly operations, and proposes a safety-aware skill augmentation framework. Distinct from object-specific policy learning, this approach augments the primitive’s compliant baseline via reinforcement learning (RL) to accommodate unmodeled contact dynamics. To ensure safe exploration, the framework integrates a Safety Policy (SP) for velocity stabilisation and employs a structured observation-action architecture. By explicitly decoupling planar motion regulation from axial compliance modulation, this architectural constraint prevents unsafe force-motion coupling during policy optimisation. Experimental validation across structurally diverse components, spanner-type configurations, and real-world products demonstrates that augmenting the TP primitive significantly enhances execution robustness and generalisability, establishing a scalable foundation for contact-rich disassembly automation. • Twist-Pull is formulated as a generic execution skill primitive for contact-rich disassembly. • A safety-aware framework augments compliant baseline skills via Reinforcement Learning. • Structured network architecture decouples motion and compliance to prevent unsafe coupling. • A dedicated Safety Policy stabilizes velocity modulation during trial-and-error exploration. • The approach generalizes to spanner-mediated operations and diverse industrial products.
Zang et al. (Fri,) studied this question.