Mulligan

Performance-Guided Data Collection for Efficient On-Robot Learning

Project · Paper · Code · Policy Arena

Release release-1: 221 LeRobot trajectory datasets, 2 simulation evaluation bundles and 180 model repositories (596 checkpoints). Every repository is MIT licensed and tagged release-1. Derived views share episodes with their raw parents, so read a card's role and parent before combining datasets.

Datasets

cNN is a collection increment, rNN a model round and bNN a recorded evaluation session. An rNN-eval dataset joins every evaluation session of one task and round on the initial state. Evaluation recordings can train later models, so check the consuming model's card before treating one as held out.

Models

596 checkpoints in 180 repositories: every checkpoint that collected, scored or was evaluated in the released datasets. Each card lists its training run config, training and validation datasets and evaluation results. .pt files are PyTorch pickles; load them only from a source you trust.

Pin the release-1 tag or its commit for reproducible reads. Generated from the release manifests of ankile/mulligan.