SEAL: Safe and Efficient Abstraction Learning for Robotic Planning

Akhil R Kurup, Ravi Prakash

Keywords: Interactive Imitation Learning, Task Planning, Semantic Modes

TL;DR: This paper introduces an active learning method to safely and efficiently teach robots the abstract subtasks needed for complex planning.

Abstract: Advances in Large Language Models have enabled the decomposition of complex tasks into subtasks for easier planning. A key challenge remains: how can we ensure a robot's task execution respects implicit task constraints and remains robust to real-world perturbations? We propose an active learning framework to efficiently learn a symbolic mode classifier that maps robot states to discrete modes, thereby establishing a grounded world model. By using active learning to iteratively improve our classification by querying experts at uncertain mode boundaries, we can match the performance of previous approaches but without any potentially unsafe perturbed trajectories or oracles, thereby enabling policies to become robust to task-level disturbances. Our findings underscore the potential for data-efficient, safe, and reliable abstraction learning to support long-term robot autonomy.

Submission Number: 8