Plan2Vec: Unsupervised Representation Learning by Latent Plans

TL;DR

Plan2vec is an unsupervised learning method that constructs a graph from image data to learn global embeddings, enabling efficient goal-conditioned control and planning over long horizons.

Contribution

It introduces a novel graph-based approach for unsupervised representation learning inspired by reinforcement learning, improving long-horizon planning and control.

Findings

Effective on simulated and real-world datasets

Achieves reactive planning with linear complexity

Successfully amortizes planning cost

Abstract

In this paper we introduce plan2vec, an unsupervised representation learning approach that is inspired by reinforcement learning. Plan2vec constructs a weighted graph on an image dataset using near-neighbor distances, and then extrapolates this local metric to a global embedding by distilling path-integral over planned path. When applied to control, plan2vec offers a way to learn goal-conditioned value estimates that are accurate over long horizons that is both compute and sample efficient. We demonstrate the effectiveness of plan2vec on one simulated and two challenging real-world image datasets. Experimental results show that plan2vec successfully amortizes the planning cost, enabling reactive planning that is linear in memory and computation complexity rather than exhaustive over the entire state space.