The dataset could not be loaded because the splits use different data file formats, which is not supported. Read more about the splits configuration. Click for more details.
Error code: FileFormatMismatchBetweenSplitsError
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
VLA Reasoning Training Dataset 1200 — Trimmed Start
Open episode 0 in the LeRobot dataset visualizer
This is a public, non-destructive derivative of
justintiensmith/VLA_Reasoning_Training_Dataset_1200 at revision
3c92bdd41891803603190b441144274544dab20b. It preserves all 1,200 episodes, prompts, actions, robot
states, and five camera streams, while omitting each episode's initial idle
prefix from the logical training sample.
Transformation
- Motion is detected independently for each episode from the six-dimensional recorded action command.
- The baseline is the component-wise median of the first 5 frames.
- Motion onset is the first run of 5 consecutive frames whose maximum absolute joint displacement from that baseline is at least 2.0 motor-position units.
- 5 frames (0.167 s) immediately before that detected onset are retained.
- Only the initial prefix is intentionally omitted by the motion heuristic.
- 14 terminal rows across 2 source episodes were also excluded to repair inherited camera/action length mismatches. This synchronization repair is reported separately from idle trimming in the manifest.
The camera files are byte-identical to the source files. Per-episode video start offsets were advanced by exactly the number of omitted frames, and the tabular rows were trimmed and reindexed to match. Thus, training and dataset visualization expose only the retained frames without introducing video re-encoding artifacts.
Result
- Episodes: 1,200
- Original logical frames: 612,733
- Retained logical frames: 570,386
- Omitted idle-prefix frames: 42,333 (6.91%)
- Source synchronization repair: 14 terminal frames across 2 episodes
- Total omitted logical frames: 42,347 (6.91%)
- Episodes with a non-zero trim: 1,176
- Frame rate: 30 FPS
The exact per-episode cut, detected onset, retained length, and prompt are in
meta/trim_audit/manifest.csv; the machine-readable configuration and checks
are in meta/trim_audit/summary.json. The source prompt manifests, environment
descriptions, and benchmark-design metadata are preserved under meta/.
Statistics note
Action and robot-state statistics were recomputed after trimming. Visual statistics are inherited from the source dataset because the retained images are byte-identical and visual inputs use model-native preprocessing rather than dataset mean/std normalization in the intended VLA-0 experiment.
- Downloads last month
- 53