Reproduce “Robotic Data Inoculation” 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
Qiyang Li, Zhiyuan Zhou, Charles Xu, Sergey Levine
The authors introduce DI, which keeps a human in control while an online-trained residual policy adds bounded, state-dependent changes to commands; the human can correct these deviations, creating recovery-relevant demonstrations. Across three real-world manipulation tasks, DI has the highest mean in all six task-by-method comparisons. For t-cube offline RL, success is 0.59 ± 0.02 versus 0.42 ± 0.06 with clean teleoperation; the findings do not establish general gains or safety.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.