Robot learning has made rapid progress on datasets of clean, successful demonstrations. But a policy trained on curated successes only knows what to do when everything goes well ... What if it doesn't?
This workshop is about other data: execution failures and near-misses, noisy or suboptimal demonstrations, out-of-distribution events, and the unexpected things people do around robots. Every lab produces it, almost none of it is documented, shared, or reused.
We want to work out, together, how to gather this data, how to curate it, and what to do with it. Can human errors provide important context, and how can we capture this? Can imperfect data help robots succeed in imperfect settings? Which failures actually matter, and how should they be labelled? What would shared infrastructure for collecting and publishing this data across academia and industry look like?
