@misc {pub6407,
	title = {CCDP: Model-free Failure Recovery via Guided Diffusion Sampling},
	author = {Amirreza Razmjoo AND Sylvain Calinon AND Michael Gienger AND Fan Zhang},
	year = {2025},
	month = {October},
	abstract = {Working in constrained environments means that failures are often inevitable, so robots must be able to recover from them. Typical recovery approaches require explicit models of the underlying task, either during learning or reproduction, to accommodate different possibilities and replan accordingly. However, such models are not always available, especially in imitation learning (IL), where one of the main advantages is precisely to avoid explicit environment/task modeling and rely instead on demonstration data. We present \textbf{CCDP (Composition of Conditional Diffusion Policies)}, a method that considers failures during inference and guides the sampling steps of diffusion policies to avoid previously failed actions. Remarkably, CCDP relies solely on successful demonstrations: it infers recovery actions without additional exploratory behavior or a high-level controller. We validate our approach on several tasks, including door opening with unknown directions, object manipulation, and button searching, and show that it consistently outperforms standard baselines.},
	publisher = {workshop paper. not included in proceedings},
	booktitle = {IROS 2025 Workshop {\textemdash} The Art of Robustness: Surviving Failures in Robotics}
}