Generalization and Failure Recovery in Motion-Planning-Based Shoelacing

Determine whether motion-planning-based robotic shoelacing systems that use predefined action primitives and designed patterns can generalize to unseen shoe and lace configurations, recover from execution failures, and enable other dexterous skills beyond the predefined behaviors.

Background

The paper discusses the limitations of classic motion-planning approaches that rely on predefined action primitives and designed patterns for the shoelacing task. Such approaches often struggle with variability in real-world conditions and the need for dexterous, precise manipulation.

The authors note that, as a consequence of these limitations, key capabilities—including generalization to unseen configurations and recovery from failures—are not yet established, motivating learning-based methods such as GR-RL to address long-horizon dexterous manipulation.