Dataset Viewer
Auto-converted to Parquet Duplicate
Search is not available for this dataset
norm_stats
dict
{ "state": { "mean": [ -0.6252827048301697, 0.05113311856985092, 0.22688305377960205, 0.46718865633010864, -0.9767469763755798, 1.7550376653671265, 3.2193922996520996, -1.5446810722351074, -2.239225149154663, 0.1927366852760315, 1.6434968709945679,...

Block Into Box PI0.5 Dataset

LeRobot v2.1 training dataset for PI0.5 LoRA fine-tuning.

Contents

  • 36 source Dataset v3 episodes recorded at 30 Hz
  • 23,060 source frames
  • 128 continuity-safe LeRobot episodes at 15 Hz
  • 11,541 converted frames
  • 256 video files and 128 Parquet files
  • 9,749 valid anchors with a complete 15-action horizon

The conversion split discontinuous source trajectories into independent episodes and dropped five segments that were too short for training. The repository also includes the validated project transforms, normalization statistics, and a SenseCore four-GPU launcher under training_bundle/.

Validation

  • all metadata, state, action, and video contracts passed
  • 384 sampled frames decoded with the expected RGB and tensor dimensions
  • all sampled numeric values were finite
  • one-step PI0.5 LoRA smoke training completed with finite loss and gradient norm

This dataset contains real robot camera recordings and is publicly accessible. Review privacy and redistribution requirements before adding new recordings.

Downloads last month
911