Grounded Robotic Action-Rule Induction through Language Models (GRAIL)

James Neidhoefer, Jacob Arkin, Nicholas Roy, Chuchu Fan

Keywords: Autonomous Planning Agents, Symbol Grounding, Frame Problem, PDDL Model Optimization, Robotics

TL;DR: GRAIL demonstrates that natural language- based symbols and symbolic systems built upon them can be derived effectively from data with little a-priori knowledge or human intervention.

Abstract: A significant body of recent work illustrates that two components of autonomous planning agents nearly always require manual pre-specification by human experts: the identification and grounding of action symbols (such as “turn right”), and the generation of PDDL action rules (including rule name, parameters, preconditions, and effects). We present the Grounded Robotic Action-Rule Induction through Language Models (GRAIL) system, which, in addition to automating those two processes, also contributes to the expanding research on PDDL model optimization. In this paper, we show how large language models (LLMs) can be used to cluster the sensorimotor experience of the robot and automatically generate useful symbolic abstractions about the robot’s capabilities and environment. This language-grounded abstraction allows the learned domain to be modified and used for planning without additional retraining. We evaluate the approach in a standard maze domain and show results for automated symbol identification and grounding, automated rule generation, simulation-based rule validation, and PDDL model optimization. We also discuss and illustrate the advantages of the hybrid neuro-symbolic GRAIL system over traditional symbolic or purely data-driven approaches to similar tasks.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 7830