Dataset Viewer
Auto-converted to Parquet Duplicate
Search is not available for this dataset
video
video
9.27
2.47k
label
class label
4 classes
0BIG
0BIG
0BIG
1MULTI
1MULTI
1MULTI
2SEP
2SEP
2SEP
2SEP
2SEP
2SEP
2SEP
2SEP
2SEP
3STACK
3STACK
3STACK
1MULTI
1MULTI
1MULTI
1MULTI
1MULTI
1MULTI
1MULTI
1MULTI
1MULTI
3STACK
3STACK
3STACK
3STACK
3STACK
3STACK
3STACK
3STACK
3STACK
1MULTI
1MULTI
1MULTI
1MULTI
1MULTI
1MULTI
1MULTI
1MULTI
1MULTI
3STACK
3STACK
3STACK
3STACK
3STACK
3STACK
3STACK
3STACK
3STACK
1MULTI
1MULTI
1MULTI
1MULTI
1MULTI
1MULTI
1MULTI
1MULTI
1MULTI
3STACK
3STACK
3STACK
3STACK
3STACK
3STACK
3STACK
3STACK
3STACK

LeRobot camera stacking benchmark videos

These are the test videos for a benchmark. The benchmark asks one question: is it faster to read robot camera views from one file per camera, or from one file that holds all the cameras?

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

Source

All frames come from lerobot/abc_130k_v3_train at revision 68651e4929d9fb00f798937b2d62617cab5c771d (Apache-2.0). We decoded its AV1 videos and encoded them again in the layouts below.

Credit for the source data goes to the ABC authors: paper and code.

There are 3 cameras: top, left_wrist and right_wrist. Each view is 224 x 224 pixels at 30 frames per second. The frames are split into 3 groups of whole episodes: g000 (8 episodes, 23,380 frames), g001 (7 episodes, 25,759 frames) and g002 (8 episodes, 24,979 frames).

Folders

Folder What it holds
variants/20260925T025221Z/ckpt_full/SEP/ One MP4 per camera per group (top, left_wrist, right_wrist). This is how LeRobot stores cameras today.
variants/20260925T025221Z/ckpt_full/STACK/ One MP4 per group. The 3 views are stacked top to bottom (top, left_wrist, right_wrist) into one 224 x 672 frame.
variants/20260925T025221Z/ckpt_full/BIG/ One MP4 per camera. Each file joins the 3 SEP groups of that camera (stream copy), so it is 3x longer. Same layout as SEP, bigger file.
variants/20260925T025221Z/ckpt_full/MULTI/ One MP4 per group with 3 video tracks, one per camera. The tracks are the SEP packets, stream copied (no new encode).
variants/r3-20260925T182847Z/ckpt_full/{STACK,MULTI}/ STACK and MULTI cut into 3 parts per group, so each part is about the size of one SEP file. Cuts start at keyframes and use stream copy (no new encode). cuts.json lists the frame ranges, bytes and sha256 of every cut.
variants/r3-20260925T180447Z-smoke/ The same cuts from the smoke run. The files are byte for byte the same as the r3 main run.
variants/r3-localtest/ckpt_1k/ Small cuts from a local test on 1,000-frame groups. Their source files are not in this repo.

Encoding

We used the LeRobot default encoder settings (VideoEncoderConfig()), set in s2_encode.py:

  • codec libsvtav1 (SVT-AV1 v3.0.0), through PyAV 15.1.0 (libavcodec 61.19.101)
  • GOP 2 (g=2): a keyframe every 2 frames
  • CRF 30
  • preset 12 (SVT-AV1 maps it to M10)
  • svtav1-params fast-decode=0
  • pixel format yuv420p, no B-frames
  • 30 fps, time base 1/30, frame i has pts i
  • MP4 with movflags faststart (index at the start of the file)

SEP and STACK files were encoded from the same decoded RGB frames. BIG, MULTI and the r3 cuts were made by stream copy, so their packets are the same bytes.

Files

Frames are per camera. A MULTI file holds 3 tracks of this many frames each.

File (under variants/20260925T025221Z/ckpt_full/) Frames MB
SEP/ckpt_full_g000__top.mp4 23,380 80.3
SEP/ckpt_full_g000__left_wrist.mp4 23,380 69.0
SEP/ckpt_full_g000__right_wrist.mp4 23,380 65.8
SEP/ckpt_full_g001__top.mp4 25,759 105.3
SEP/ckpt_full_g001__left_wrist.mp4 25,759 81.9
SEP/ckpt_full_g001__right_wrist.mp4 25,759 77.7
SEP/ckpt_full_g002__top.mp4 24,979 67.9
SEP/ckpt_full_g002__left_wrist.mp4 24,979 56.5
SEP/ckpt_full_g002__right_wrist.mp4 24,979 56.7
STACK/ckpt_full_g000.mp4 23,380 211.5
STACK/ckpt_full_g001.mp4 25,759 260.6
STACK/ckpt_full_g002.mp4 24,979 177.2
BIG/ckpt_full_b000__top.mp4 74,118 253.5
BIG/ckpt_full_b000__left_wrist.mp4 74,118 207.4
BIG/ckpt_full_b000__right_wrist.mp4 74,118 200.1
MULTI/ckpt_full_g000.mp4 23,380 215.3
MULTI/ckpt_full_g001.mp4 25,759 265.2
MULTI/ckpt_full_g002.mp4 24,979 181.4

r3 cuts (variants/r3-20260925T182847Z/ckpt_full/). STACK and MULTI use the same frame ranges.

Cut Frames in group Frames STACK MB MULTI MB
g000_p0 0 to 8,266 8,266 70.5 71.9
g000_p1 8,266 to 14,066 5,800 70.5 71.6
g000_p2 14,066 to 23,380 9,314 70.5 71.9
g001_p0 0 to 6,626 6,626 86.8 88.1
g001_p1 6,626 to 17,092 10,466 86.9 88.6
g001_p2 17,092 to 25,759 8,667 86.8 88.5
g002_p0 0 to 7,336 7,336 59.1 60.3
g002_p1 7,336 to 14,600 7,264 59.1 60.3
g002_p2 14,600 to 24,979 10,379 59.1 60.8

In total there are 75 data files (72 MP4 and 3 cuts.json), 5.29 GB.

Related datasets

Small LeRobot v3 datasets built on these videos, used by the streaming and PR #3917 tests:

License

Apache-2.0, the same as the source dataset.

Downloads last month
36

Paper for hadriencornier/lerobot-camera-stacking-benchmark