Learning Task Specifications from Demonstrations as Probabilistic Automata

TL;DR

This paper presents a novel, efficient method for learning probabilistic automata from demonstrations, enabling robots to understand and adapt to complex tasks with interpretable structures.

Contribution

It introduces a new approach for inferring task structures and preferences from demonstrations using probabilistic automata, improving interpretability and adaptability.

Findings

Successfully learned task structures from demonstrations

Enabled robots to replicate diverse expert strategies

Demonstrated adaptability to changing conditions

Abstract

Specifying tasks for robotic systems traditionally requires coding expertise, deep domain knowledge, and significant time investment. While learning from demonstration offers a promising alternative, existing methods often struggle with tasks of longer horizons. To address this limitation, we introduce a computationally efficient approach for learning probabilistic deterministic finite automata (PDFA) that capture task structures and expert preferences directly from demonstrations. Our approach infers sub-goals and their temporal dependencies, producing an interpretable task specification that domain experts can easily understand and adjust. We validate our method through experiments involving object manipulation tasks, showcasing how our method enables a robot arm to effectively replicate diverse expert strategies while adapting to changing conditions.