An Unbiased Look at Datasets for Visuo-Motor Pre-Training

TL;DR

This paper emphasizes the importance of dataset choice over algorithms in visuo-motor pre-training for robotics, revealing that traditional vision datasets are surprisingly effective and that dataset distribution impacts learning more than size.

Contribution

It shifts focus from algorithm development to dataset analysis, showing traditional vision datasets are competitive and dataset distribution influences transfer success more than dataset size.

Findings

Traditional vision datasets are surprisingly effective for visuo-motor learning.

Dataset image distribution impacts transfer performance more than dataset size.

Simple regularization improves real-world policy learning.

Abstract

Visual representation learning hold great promise for robotics, but is severely hampered by the scarcity and homogeneity of robotics datasets. Recent works address this problem by pre-training visual representations on large-scale but out-of-domain data (e.g., videos of egocentric interactions) and then transferring them to target robotics tasks. While the field is heavily focused on developing better pre-training algorithms, we find that dataset choice is just as important to this paradigm's success. After all, the representation can only learn the structures or priors present in the pre-training dataset. To this end, we flip the focus on algorithms, and instead conduct a dataset centric analysis of robotic pre-training. Our findings call into question some common wisdom in the field. We observe that traditional vision datasets (like ImageNet, Kinetics and 100 Days of Hands) are surprisingly competitive options for visuo-motor representation learning, and that the pre-training dataset's image distribution matters more than its size. Finally, we show that common simulation benchmarks are not a reliable proxy for real world performance and that simple regularization strategies can dramatically improve real world policy learning. https://data4robotics.github.io