Skip to content

Dataset schema — a superset that stays LeRobot-safe·

The first 10 dims of observation.state match the official earthrover_mini_plus / lilkm/earthrover-navigation layout (same dotted names, same order) so policies and checkpoints transfer — a standard model slices [:10]. scout then appends richer raw telemetry.

observation.state — 27-D·

idx fields source
0–9 linear.vel · angular.vel · battery.level · orientation.deg · gps.{lat,lng,signal} · signal.level · vibration · lamp.state official core (transfer-compatible)
10–11 voltage · current electrical
12–17 imu.{accel,gyro}.{x,y,z} latest IMU sample
18–20 imu.mag.{x,y,z} magnetometer — absolute heading
21–24 rpm.{fl,fr,rl,rr} per-wheel odometry — ground-truth proprioception (commanded ≠ executed)
25–26 power · network_state system

action — 2-D, normalized [-1, 1]·

linear.vel · angular.vel — the reference formulation; the lamp lives in state.

Video & audio·

observation.images.front and observation.images.rear (rear on by default; ROVER_RECORD_REAR=0 for front-only) at ROVER_RECORD_FPS (10) and ROVER_RECORD_WIDTH×HEIGHT (640×480), encoded SCOUT_VCODEC (h264); observation.audio at ROVER_AUDIO_RATE (16 kHz) from the hub's mic stream.

Burst arrays

The SDK exposes IMU/mag/rpm as bursts (several samples per poll with a unix timestamp). scout takes the freshest sample per 10 Hz frame.

Sidecars, aligned by frame index·

datasets/scout__earth-rover-mini-YYYYMMDD/<persona>/
├── meta/info.json  meta/episodes/…  meta/tasks.parquet
├── data/chunk-000/file-000.parquet            # one file per sealed episode
├── videos/observation.images.front/chunk-000/file-000.mp4
├── reasoning/events.sqlite                    # ECoT: prompt → tool calls → results, frame_index = round((ts - t0) * fps)
└── detections/episode_000000.jsonl            # YOLO per frame: {frame, ts, cam, dets:[{cls, conf, xyxy}]}
  • ECoT (tools/reasoning_log.py, tools/ecot_export.py) binds every reasoning step to the video spine; the export to 🤗 cagataydev/scout-earthrover-ecot flattens it.
  • Detections (yolo_detector.py) are keyed by the recorder's frame index — same captured frame, no drift.
  • Merging — tools/merge_datasets.py folds per-persona datasets into a daily corpus; tools/dataset_index.py lists them for the cockpit and the agent prompt (SCOUT_INJECT_DATASETS).

Why one parquet + one mp4 per episode, and how to repair an index: Datasets & replay.