Alan0928/robot-agent-gallery

Robot Evaluation Gallery Media

Browse and play the gallery · Read the essay

This public dataset stores the complete media and evidence archive for the companion static Hugging Face Space. Keeping media in a Dataset avoids the Space repository's 1 GB storage limit. The Space contains the gallery interface and reads videos directly from this repository.

The current gallery presents 1,286 scored evaluation episodes from 497 task entries across LIBERO, Robotwin, RoboCasa365 and RoboDojo. The archive also retains the historical 50 Robotwin episodes, 490 training demonstrations (including historical references), and nine additional diagnostic rerun recordings. The complete media archive contains 1,835 MP4s and 1,835 JPEG posters, totaling 4,082,694,506 bytes.

data/gallery.json and data/episodes.csv preserve the complete historical evaluation archive (1,336 episodes, 547 task entries). data/task-demos.json records training demonstrations and explicit unavailable states. Other JSON and CSV files preserve outcome, selection, provenance and content hashes. Media paths remain relative, under media/. Historical RoboCasa Release URLs remain in metadata as provenance; current playback loads from this Dataset.

The original export and validation tools live in the companion Space source. Download this Dataset's media/ into a local Space checkout to run the full offline media audit. Follow METHODOLOGY.md, TRAINING_DEMOS.md and import notes for source selection and interpretation limits. These documents include historical publication procedures.

Success is each benchmark's native score. Training demonstrations and diagnostic reruns do not contribute to gallery evaluation totals. Different task distributions, controls and budgets prevent pooling success rates.