Comparing Reconstruction- and Contrastive-based Models for Visual Task Planning

TL;DR

This paper compares reconstruction-based and contrastive-based models for visual task planning, showing that models using task priors like Siamese networks with contrastive loss outperform traditional reconstruction methods in learning effective state representations from images.

Contribution

The study provides a comprehensive evaluation of different loss functions for state representation learning, highlighting the superiority of contrastive models with task priors for visual planning.

Findings

Contrastive models outperform reconstruction-based models in visual task planning.

Task priors like Siamese networks improve state representation quality.

Evaluation metrics tailored for task relevance are essential for assessing models.

Abstract

Learning state representations enables robotic planning directly from raw observations such as images. Most methods learn state representations by utilizing losses based on the reconstruction of the raw observations from a lower-dimensional latent space. The similarity between observations in the space of images is often assumed and used as a proxy for estimating similarity between the underlying states of the system. However, observations commonly contain task-irrelevant factors of variation which are nonetheless important for reconstruction, such as varying lighting and different camera viewpoints. In this work, we define relevant evaluation metrics and perform a thorough study of different loss functions for state representation learning. We show that models exploiting task priors, such as Siamese networks with a simple contrastive loss, outperform reconstruction-based representations in visual task planning.