Explainable and Interpretable Methods for Handling Robot Task Failures

Doctoral thesis, 2025

Robots are increasingly deployed in dynamic human environments. To avoid failures during task execution, such as setting a table, they must adapt to unexpected changes. This is challenging because robots must proactively predict failures and identify controllable factors to prevent them. If failures cannot be autonomously prevented, robots should explain failure causes, which is challenging because explanations should cater to non-expert users. The first goal of this thesis is therefore to enhance the reliability and explainability of robots by predicting, explaining, and preventing task execution failures using symbolic causal models. We introduce a novel framework for learning causal models from simulated data. To improve the transferability of the causal models between tasks, we propose three parameter transfer methods that leverage the semantic similarities between models. To enhance failure prediction, we propose a novel approach that combines the learned causal models with a breadth-first search procedure for proactive failure prediction and contrastive failure explanation. We validate this approach on object manipulation tasks, such as stacking cubes, achieving a 95% failure prevention rate. We then extend the method to predict human perceptions of a navigation robot's competence and improve its behavior, resulting in a 72% increase in perceived competence.

Another common failure is missing capabilities that hinder a robot from achieving its task goal. The second thesis goal is therefore to enable non-experts to assist by teaching robots the missing actions intuitively, without coding experience. We propose a novel demonstration system that lets users teach tasks in Virtual Reality. Our system automatically segments and classifies the demonstrations, generating symbolic, robot-agnostic actions that integrate into the robot's existing capabilities. Our approach achieves a 92% success rate in learning task abstractions from a single demonstration in single- and multi-agent tasks. Additionally, our approach enables robots to detect missing actions automatically, allowing users to demonstrate only the missing parts instead of the entire task, reducing demonstration time by 61%.

The presented contributions enable robots to handle dynamic environments more reliably and explainably while continuously expanding their capabilities to adapt to new challenges.