Learning Neuro-Symbolic Relational Transition Models for Bilevel Planning

TL;DR

This paper introduces Neuro-Symbolic Relational Transition Models (NSRTs), which combine symbolic and neural methods to enable efficient learning and planning in complex robotic domains with continuous states, actions, and long horizons.

Contribution

The work presents a novel neuro-symbolic model that is data-efficient, generalizes over objects, and integrates symbolic AI planning with neural models for robotic planning.

Findings

NSRTs learn within hundreds of episodes.

They enable fast planning for tasks with up to 60 actions.

They generalize to new objects and larger tasks.

Abstract

In robotic domains, learning and planning are complicated by continuous state spaces, continuous action spaces, and long task horizons. In this work, we address these challenges with Neuro-Symbolic Relational Transition Models (NSRTs), a novel class of models that are data-efficient to learn, compatible with powerful robotic planning methods, and generalizable over objects. NSRTs have both symbolic and neural components, enabling a bilevel planning scheme where symbolic AI planning in an outer loop guides continuous planning with neural models in an inner loop. Experiments in four robotic planning domains show that NSRTs can be learned after only tens or hundreds of training episodes, and then used for fast planning in new tasks that require up to 60 actions and involve many more objects than were seen during training. Video: https://tinyurl.com/chitnis-nsrts