Reproduce “RoboNet: Large-Scale Multi-Robot Learning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Sudeep Dasari, Frederik Ebert, Stephen Tian, Suraj Nair, Bernadette Bucher, Karl Schmeckpeper, Siddharth Singh, Sergey Levine and 1 more
Researchers from UC Berkeley, Stanford University, and collaborators present RoboNet, a large-scale, multi-robot dataset for learning generalizable vision-based manipulation skills. They show that pre-training on RoboNet enables models to achieve improved zero-shot generalization to novel viewpoints and backgrounds, and facilitates few-shot adaptation to new robot hardware, outperforming methods trained with up to 20 times more robot-specific data.
Key insights
- Large-scale, diverse real-world datasets like RoboNet can enable broad generalization in robotic learning, allowing models to transfer skills across varied hardware and environments.
- Pre-training on RoboNet significantly reduces the data required for adapting to new robot platforms, outperforming robot-specific training with 4x-20x more data in some cases.
- Current visual prediction models, even with hundreds of millions of parameters, exhibit underfitting on RoboNet, indicating a need for higher capacity models or more intelligent data selection strategies.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.