Part 1-2

Forward & Backward Skill Compatibility

AblationPolicy only cannot use the new skill. Interface + random policy still fails from inconsistent subtask selection.

Full systemThe learned policy proposes a coherent sequence, while validation + hooking remaps the mismatched skill and completes the long-horizon task.

SIL-C Figure 5 ablation comparing policy only, SIL-C with a random policy, and SIL-C with a learned policy and interface — SIL-C · Figure 5