Plan2Vec: Unsupervised Representation Learning by Latent Plans

Ge Yang,^† Amy Zhang,^† Ari S. Morcos,^† Joelle Pineau,^†§ Pieter Abbeel,^‡ Roberto Calandra^†

^*Equal Contribution, ^†Facebook AI Research, ^§McGill University, ^‡UC Berkeley

CODE|PAPER|SLIDES

Overview

Plan2vec is an unsupervised representation learning method that uses graph-search to learn long-horizon relationships between images.

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.

BibTex

@inproceedings{yang2020plan2vec,
    title={Plan2vec: Unsupervised Representation Learning by Latent Plans},
    author={Yang, Ge and Zhang, Amy and Morcos, Ari S. and Pineau, Joelle
            and Abbeel, Pieter and Calandra, Roberto},
    booktitle={Proceedings of The 2nd Annual Conference on Learning for Dynamics and Control},
    series={Proceedings of Machine Learning Research},
    pages={1-12},
    year={2020},
    volume={120},
    note={arXiv:2005.03648}
}