Reproduce “Mulligan: Performance-Guided Data Collection for Efficient On-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
Lars Ankile, Perry Dong, Rohan Bhowmik, Aneesh Muppidi, David D. Yuan, Shuran Song, Chelsea Finn
Mulligan guides robot-learning data collection by choosing some starting states where the policy failed while using coverage-based choices to explore less-tried states. Operators correct and replay tasks, while a critic learns from all rollouts. On Insert Marker, full HiL-IDQL+Mulligan achieved 76% success versus 42% for uniform HG-DAgger; evidence is limited to studied tasks and operator-placeable starts.
Key insights
- Episode selection is itself an important lever for data efficiency: observed failures can guide supervision toward states where the policy still needs help.
- Combining targeted retries with coverage helps address known weaknesses without ignoring untested states; counterfactual demonstrations and critic reranking provide complementary benefits at different learning stages.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.