A Framework for Neurosymbolic Robot Action Planning using Large Language Models

TL;DR

Teriyaki introduces a neurosymbolic planning framework that uses GPT-3 fine-tuned on domain-specific problem-plan pairs to produce PDDL-compatible plans in a streaming, action-by-action fashion. It addresses the scalability and fluency limitations of symbolic planners in human-robot collaboration by enabling concurrent planning and execution. Experiments in two articulated-object manipulation domains show planning validity comparable to a state-of-the-art PDDL planner, shorter plans by up to 13.5%, and a reduction in plan availability waiting time by up to 61.4%. While not yet a complete replacement for traditional planners, Teriyaki demonstrates a viable data-driven approach to domain-specific planning with clear benefits for reactive, fluent robotic systems.

Abstract

Symbolic task planning is a widely used approach to enforce robot autonomy due to its ease of understanding and deployment in robot architectures. However, techniques for symbolic task planning are difficult to scale in real-world, human-robot collaboration scenarios because of the poor performance in complex planning domains or when frequent re-planning is needed. We present a framework, Teriyaki, specifically aimed at bridging the gap between symbolic task planning and machine learning approaches. The rationale is training Large Language Models (LLMs), namely GPT-3, into a neurosymbolic task planner compatible with the Planning Domain Definition Language (PDDL), and then leveraging its generative capabilities to overcome a number of limitations inherent to symbolic task planners. Potential benefits include (i) a better scalability in so far as the planning domain complexity increases, since LLMs' response time linearly scales with the combined length of the input and the output, and (ii) the ability to synthesize a plan action-by-action instead of end-to-end, making each action available for execution as soon as it is generated instead of waiting for the whole plan to be available, which in turn enables concurrent planning and execution. Recently, significant efforts have been devoted by the research community to evaluate the cognitive capabilities of LLMs, with alternate successes. Instead, with Teriyaki we aim to provide an overall planning performance comparable to traditional planners in specific planning domains, while leveraging LLMs capabilities to build a look-ahead predictive planning model. Preliminary results in selected domains show that our method can: (i) solve 95.5% of problems in a test data set of 1,000 samples; (ii) produce plans up to 13.5% shorter than a traditional symbolic planner; (iii) reduce average overall waiting times for a plan availability by up to 61.4%

Figures (7)

• Figure 1: A Baxter robot executing actions in two domains involving the manipulation of an articulated object. A human can act on an articulated object's joint at any time, forcing the robot to re-plan.
• Figure 2: A diagram of Teriyaki fine-tuning process. Blocks in yellow represent custom code developed for data generation and processing as described in Section \ref{['dataset']}.
• Figure 3: A possible architecture integrating a generative Teriyaki planner to implement simultaneous planning and execution. The dashed line below the planner represents the fact that actions are generated one by one, and then buffered waiting for their execution. The SPEM module checks that each action's preconditions are satisfied before submitting it for execution.
• Figure 4: Evolution of the token validation accuracy and the planning validity as the number of training examples increases. The blue line represents the evolution of the validation token accuracy during learning as reported by GPT-3 classification metrics. The bars represent the planning validity of the Teriyaki-MACRO, Teriyaki-NO-MACRO (davinci) and Teriyaki-NO-MACRO (MACRO) models. The red line represents our baseline, that is, the percentage of plans solved by Probe.
• Figure 5: A diagram of Teriyaki testing process as described in Section \ref{['sub:solvers']}.