Robot skills that work independently often break when composed into a longer task. The new paper identifies three recurring failure modes: a tool or skill may lack a required capability, a completion signal may not reflect true physical progress, and one skill may leave the robot in a state that violates the prerequisites of the next skill.
The authors propose a framework that builds shared executable stacks and then recursively improves them, allowing the robot to compose skills for long-horizon tasks without prior examples. The emphasis is on zero-shot generalization: the system should handle new task sequences by reusing and refining these stacks rather than learning each combination from scratch.
Because this is a preprint on arXiv, the work has not yet undergone peer review. The abstract outlines the problem and the proposed direction, but the full details of the recursive improvement mechanism and experimental results would require reading the paper itself.