ego-steer/EgoSteer-example-data

EgoSteer example data

Anonymous example subsets. This repository contains small VLA and VLM samples for setup verification, not the complete EgoSteer training datasets. The LeRobot anonymous review subsets are RealWorld-subset (1,920 episodes across 160 tasks) and Egocentric-subset (40 episodes from 8 sources).

A small WebDataset package for verifying an EgoSteer setup end to end: compute the state/action normalizer, launch single-node training, and run offline evaluation. It is not the full training set; its small sample shards let you exercise the pipeline without preparing your own data.

Layout

vla/
  train/shard-*.tar   # 10 shards, dual-camera (head + chest) real-robot frames, 30 fps, per-episode contiguous
  val/shard-*.tar     # 1 shard
vlm/
  train/shard-*.tar   # image + text QA samples (from public VLM datasets)
  val/shard-*.tar

Each VLA frame is a WebDataset sample with image.jpg, depth.npy, chest_image.jpg, chest_depth.npy, lowdim.npy (136-D: wrist/hand state and action + head/chest camera extrinsics and intrinsics) and meta.json (instruction, instruction_num, episode_index, dataset_name, cameras, presence). The exact field layout and coordinate conventions are documented in data/data.md of the code repository.

Download

hf download ego-steer/EgoSteer-example-data --repo-type dataset --local-dir example_data

The code repository's default configs point at example_data/ under the repository root.

License

Released under the Apache 2.0 license.