Reproduce “Learning Robot Behavior Priors from Non-Robotic Synthetic Data” 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
Mateusz Odrowaz-Sypniewski, Michael Amir, Jasmine Bayrooti, Ajay Shankar, Amanda Prorok
The authors train a Transformer on generated control tasks without robot data, then predict actions on unseen robot datasets. At evaluation, examples from each task provide context. Across four datasets, control-prior pretraining reaches mean accuracy of 0.166 ± 0.000, versus 0.039 ± 0.009 for a generic tabular prior. This measures action prediction, not successful robot control.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.