ReSIP: Reinforcement Learning with Symbolic Inductive Planning for Interpretable and Generalizable Pixel-Based Control

ICLR 2026 Conference Submission12693 Authors

Keywords: Planning and Learning, Symbolic Regression, Compositional Generalization, Object-Centric Representation

TL;DR: We propose ReSIP, a novel framework for pixel-based control that combines Reinforcement Learning with Symbolic Inductive Planning, aiming to deal with long-horizon sequential task and object manipulation.

Abstract: Deep Reinforcement Learning (DRL) has struggled with pixel-based controlling tasks that have long sequences and logical dependencies. Methods using structured representations have shown promise in generalizing to different objects in manipulation tasks. However, they lack the ability to segment and reuse atomic skills. Neuro-symbolic RL excels in handling long sequential decomposable tasks yet heavily relies on expert-designed predicates. To address these challenges, we propose ReSIP, a novel framework for pixel-based control that combines Reinforcement Learning with Symbolic Inductive Planning. Our approach first automatically discovers and learns atomic skills through experiences in simple environments without human intervention. Then, we employ a genetic algorithm to enhance these atomic skills with symbolic interpretations. Therefore, we convert the complex controlling problem into a planning problem. Taking advantage of symbolic planning and object-centric skills, our model is inherently interpretable and provides compositional generalizability. The results of the experiments show that our method demonstrates superior performance in long-horizon sequential tasks and complex object manipulation.

Primary Area: interpretability and explainable AI

Submission Number: 12693