An Unbiased Look at Datasets for Visuo-Motor Pre-Training
This work addresses dataset selection for robotics pre-training, revealing insights that challenge common assumptions, though it is incremental as it focuses on analysis rather than new methods.
The study tackled the problem of visual representation learning for robotics by analyzing the impact of pre-training dataset choice, finding that traditional vision datasets are competitive and image distribution matters more than size, and showing that simulation benchmarks are unreliable for real-world performance while simple regularization improves policy learning.
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