Skip to content
Bud's Wiki

Shotwell releases 415+ hours of annotations on ABC-130K

TL;DR

Section titled “TL;DR”

Shotwell (@ShotwellSt, tagline “Make the most of your robotics data”) announces the release of 415+ hours of annotations layered on top of the ABC-130K bimanual-manipulation dataset from XDOF (ABC: Scalable Behavior Cloning with Open Data, Training, and Evaluation). The team ran five tasks from ABC-130K through their “Shotwell engine” and posted results at demo.shotwell.ai. The tweet is an artifact-thin announcement — the concrete asset is the annotations bundle sitting behind the demo site, not a paper or model.

Key claims

Section titled “Key claims”
  • Shotwell releases 415+ hours of annotations on top of XDOF’s ABC-130K dataset [tweet].
  • Five ABC-130K tasks were processed through the “Shotwell engine” to produce these annotations [tweet].
  • Results are viewable at demo.shotwell.ai (with a padlocked variant at demo.shotwell.ai/demo-padlock) [tweet].

Method

Section titled “Method”

Not disclosed in the tweet. Shotwell markets itself as a robotics-data annotation service (“Make the most of your robotics data”); the “engine” appears to be their internal annotation pipeline applied to teleoperation episodes. The demo site is the primary artifact — no dataset dump on Hugging Face, no code, no technical write-up is linked from the announcement. A follow-up reply from @abidlabs asks whether the annotations will be released as a Hugging Face dataset; no answer is given in the visible thread.

Results

Section titled “Results”
  • 415+ hours of annotations covering 5 ABC-130K tasks, viewable on the demo site [tweet].
  • No accuracy numbers, comparison to human labelers, or downstream policy-learning metrics reported in the tweet.

Why it’s interesting

Section titled “Why it’s interesting”

Layers a new labeling pass on top of the largest filed open bimanual teleoperation dataset (ABC: Scalable Behavior Cloning with Open Data, Training, and Evaluation ships 3,553h / 134,806 episodes / 195 tasks unlabeled beyond ABC’s own primitive categories) — if the annotations are useful for language-conditioning or subtask segmentation, they plug directly into the same VLA-recipe hole that Segmenting Robot Video into Actionable Subtasks (WGO-Bench) (WGO-Bench) and Introducing Perceptron Egocentric API target from the egocentric side. Complementary to the “annotate other people’s manipulation data” trend also visible in Pareto (Hebbian Robotics) indexes AllenAI's MolmoAct2 Bimanual YAM dataset — clusters + prompt normalization (Brandon Ong) (Pareto indexing MolmoAct2 Bimanual YAM) and TrueLabel — Physical AI Dataset Directory: Robotics, Humanoid & Egocentric Data (TrueLabel’s dataset directory) — the labor of cataloging and enriching released robotics datasets is emerging as a distinct commercial layer.

See also

Section titled “See also”