hadriencornier/lerobot-data11-r3-stack

DATA-11 round 3 test dataset, STACK layout

A small LeRobot v3 dataset for speed tests. It is part of a benchmark of stacking camera views, used by the PR #3917 tests.

Code and results: https://github.com/Hadrien-Cornier/lerobot-camera-stacking-benchmark

Test videos: hadriencornier/lerobot-camera-stacking-benchmark

What is inside

• 232 episodes, 592,944 frames, 30 fps.
• Cameras: one video key, observation.images.stack: the 3 views (top, left_wrist, right_wrist) stacked top to bottom in one 224 x 672 frame.
• 72 video files: 8 copies of the 9 r3 STACK cuts from the benchmark repo. Each cut is about the size of one SEP file. An episode that crosses a cut is split in two, so this dataset and lerobot-data11-r3-sep have the same episodes.
• observation.state and action are random float32 values (14 each). The task text is a placeholder (pick and place). Use this dataset to test reading speed, not to train a policy.
• Videos: AV1 (libsvtav1), GOP 2, CRF 30, preset 12, yuv420p. The same bytes as the files in the test video repo.

Source

The video frames come from lerobot/abc_130k_v3_train at revision 68651e4929d9fb00f798937b2d62617cab5c771d (Apache-2.0). Credit for the source data goes to the ABC authors: paper and code.

License

Apache-2.0, the same as the source dataset.