Language Guided Operator Learning for Goal Inference
Keywords: Goal Inference, Operator Learning, Language Models
TL;DR: We propose a new method for operator learning that combines symbolic approaches with LLMs to discover additional preconditions, and we show how the learned operator library can help predict the goals of other agents from demonstrations.
Abstract: Accurately predicting the goals of other agents is an essential skill for intelligent agents deployed in interactive environments. However, existing methods for goal inference struggle in scenarios with open-ended goal spaces or with complex action sequences. In this work, we present a novel approach for online goal inference based on library learning and explanation-based inference, guided by large language models. Our algorithm learns a library of operators from demonstrations using symbolic learning methods and guided by large language models. At inference, it predicts other agents' goals from partial plans by explaining the agents' past actions based on the learned library of operators. Specifically, we introduce an algorithm called precondition parsing which analyzes observed action sequences and hypothesizes future actions based on the relationship between observed actions. We evaluate our approach in a 2D Minecraft-like domain and show that both the library learning and explanation-based prediction significantly improve the ability to predict the goals of other agents.
Submission Number: 29
Loading