Leveraging Temporal Structure in Safety-Critical Task Specifications for POMDP Planning

Jason Xinyu Liu, Eric Rosen, Suchen Zheng, Stefanie Tellex, George Konidaris

Abstract: Navigating a partially observable environment while satisfying temporal and spatial constraints is an essential safety feature of many robotic applications. For example, an autonomous drone needs to understand the command “Find the supermarket while avoiding the park” to avoid possible collisions with trees. Previous approaches chose to sacrifice generality for computational efficiency in large state spaces by designing action heuristics that do not apply across different tasks or used a value-iteration-based planner that does not scale well. Our approach automatically extracts structured rewards from linear temporal logic (LTL) task specifications to guide a sampling based POMDP planner, named LTL-POMCP. We augment a partially observable Markov decision process (POMDP) with an LTL task specification then use LTL-POMCP to solve the resultant composite POMDP. Quantitative results from a classic POMDP domain show that LTL-POMCP can generalize to various LTL task specifications and scale to large state spaces. We then demonstrate the first end-to-end system from temporally constrained natural language to robot policies in partially observable maps in simulation.