Mulligan

Performance-Guided Data Collection for Efficient On-Robot Learning

Project and paper · Policy Arena · Full dataset catalog and mapping

September 22, 2026 release, with evaluations regrouped into one dataset per task-round on September 24: 241 verified trajectory repositories plus two simulation evaluation bundles. The Arena selects mainline results and required training ancestry. Derived views share episodes with their raw parents.

Insert Marker — 50 trajectory datasets
Thread Nut — 53 trajectory datasets
Route Cable — 72 trajectory datasets
Square-Narrow — 34 trajectory datasets
Square-Broad — 37 trajectory datasets

cNN: collection increment · rNN: model round · bNN: evaluation block. Consult the card and training recipe before mixing views or treating a historical evaluation as held out.

Source history is preserved in source-main. Use pinned release commits/tags for reproducibility. Historical standalone D1/pilot campaigns are excluded; D1-named Routing sources are retained only as required D2 ancestry.