Planning with Goal-Conditioned Policies

TL;DR

This paper introduces a method combining goal-conditioned reinforcement learning with planning, using latent variable models to create state abstractions from complex observations, enabling effective multi-stage decision making in robotics.

Contribution

It proposes a novel approach that integrates RL-learned goal-conditioned policies with planning, utilizing latent variable models to abstract states from high-dimensional inputs like images.

Findings

Outperforms prior methods on image-based robot navigation tasks.

Enables planning with learned policies for complex, multi-stage behaviors.

Provides a scalable way to abstract states from high-dimensional observations.

Abstract

Planning methods can solve temporally extended sequential decision making problems by composing simple behaviors. However, planning requires suitable abstractions for the states and transitions, which typically need to be designed by hand. In contrast, model-free reinforcement learning (RL) can acquire behaviors from low-level inputs directly, but often struggles with temporally extended tasks. Can we utilize reinforcement learning to automatically form the abstractions needed for planning, thus obtaining the best of both approaches? We show that goal-conditioned policies learned with RL can be incorporated into planning, so that a planner can focus on which states to reach, rather than how those states are reached. However, with complex state observations such as images, not all inputs represent valid states. We therefore also propose using a latent variable model to compactly represent the set of valid states for the planner, so that the policies provide an abstraction of actions, and the latent variable model provides an abstraction of states. We compare our method with planning-based and model-free methods and find that our method significantly outperforms prior work when evaluated on image-based robot navigation and manipulation tasks that require non-greedy, multi-staged behavior.