LOOP: A Plug-and-Play Neuro-Symbolic Framework for Enhancing Planning in Autonomous Systems

ICLR 2026 Conference Submission13564 Authors

Keywords: Planning with Language Models, Neuro-Symbolic Learning, Causal Learning, Learning for Planning and Scheduling, Coordination and Collboration

TL;DR: LOOP is a closed-loop neuro-symbolic planner that iteratively synthesizes/refines PDDL and learns a causal knowledge base from executions; 85.8% success on six IPC domains, outperforming prior LLM-based planners.

Abstract: Planning is one of the most critical tasks in autonomous systems, where even a minor error can lead to significant failures or losses. Current state-of-the-art neural planners struggle in complex domains, often producing plans with missing preconditions, inconsistent goals, or hallucinated steps, while classical planners provide guarantees but lack the flexibility and natural-language understanding needed in modern systems. Existing neuro-symbolic methods typically perform a one-shot translation from natural language to formal plans. In safety-critical autonomous systems, this leaves no mechanism to detect and correct specification errors before execution. To address this, we introduce LOOP, a neuro-symbolic planning framework that models planning as an iterative interaction between neural and symbolic modules. It synthesizes Planning Domain Definition Language (PDDL) models from task descriptions, refines them using feedback from a symbolic planner and execution rollouts, and builds a causal knowledge base from traces to guide subsequent plans. Across six International Planning Competition (IPC) domains, LOOP attains 85.8% task success, surpassing LLM+P (55.0%), LLM-as-Planner (19.2%), and Tree-of-Thoughts (3.3%). Together, these results indicate that consistent planning arises from sustained interaction between neural and symbolic reasoning rather than one-shot translation.

Primary Area: neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)

Submission Number: 13564