itaimizlish/daniel-yard-depth

DANIEL yard depth: benchmark, training set and hardware-in-the-loop logs

The data behind Seeing to Drive on a Single Board (Itai M., 2026): a distilled depth network on a Raspberry Pi 5 + Hailo-8L driving a Traxxas through a Gaussian-splat reconstruction of a real yard. Project page · paper · model

Contents

Folder	Size	What
bench/	1.8 GB	The depth benchmark: 400 held-out simulated frames with the renderer's metric depth, the car's pose and camera intrinsics, 100 held-out real robot frames, and every evaluated model's predictions (7 Hailo-8L builds read back from the chip, 10 float GPU models)
train/	9.7 GB	The student's training set as tar shards: 19,491 RGB frames (img-*.tar), rendered-GT/teacher mix labels (labels_gtmix-*.tar) and Depth Anything V2-Large teacher labels (labels_vitl-*.tar), both 280×504 float16 disparity
hitl/	70 MB	Per-step logs of every hardware-in-the-loop arm (runs.tar.zst: pose, commands, Pi timing, SLAM traces, per-arm summaries), the scenes, the prop footprints and the scored arm table

bench/

• frames.json: one entry per simulated frame: img, gt (metric depth in metres, float16 280×504, inf = sky / no hit), scene (07 crates, 08 people and cones, 09 barrels: never seen in training), run, step, pose (x, y, yaw of the car in the yard frame) and meta (camera fx, fy, cx, cy, height).
• sim/b000.jpg … b399.jpg (+ _depth.npy), real/r000.jpg … r099.jpg.
• predictions/: preds_<model>.npy / real_<model>.npy (GPU, float16), out_<hef>.npz (on-chip outputs with each model's crop box and output kind).
• scores.json and timing files: the numbers in the paper's Table I, produced by the project's scoring script (code available on request).
• scenarios.json, props_manifest.json: prop placements and collider footprints, used to mark obstacle pixels.

train/

frames.json lists every id with its split and kind: 13,902 train + 3,465 val simulated, 2,024 train + 100 val real; frames_v1.json and frames_v3.json are the subsets used by students v1 and v3/v5. Untar into one directory to get img/<id>.jpg, labels_gtmix/<id>.npy, labels_vitl/<id>.npy.

hitl/

runs.tar.zst unpacks to runs/<arm>_pi/<scene>.jsonl (one JSON line per control step: pose, goal, steer, throttle, mode, clearance, round trip and Pi stage times), <scene>_slam.json (own-pose runs) and summary_<arm>_pi.json. hitl_arms.json holds the per-seed clean/reached scores on the car's real 0.557×0.294 m footprint, as reported.

How it was made

Simulated frames are rendered by Isaac Sim from a NuRec/3DGUT Gaussian-splat reconstruction of one real yard, with eleven kinds of public 3D props (chairs, a cone, a box, a barrel, a bin, a sign, a crate, two people) placed on the reconstructed ground. The renderer's depth has speckle and floaters near the floor, so the benchmark also reports an obstacle region: pixels that, back-projected with the true pose, land on a prop's collider. Real frames come from the robot's own camera in a different building and have no ground truth.

Not included

The yard reconstruction itself (a private property) and the prop USD assets (their own licences). The frame dumps of every episode (138 GB) are not released; the per-step logs are.

Licence and citation

CC-BY-NC-4.0 (the teacher labels come from Depth Anything V2-Large, CC-BY-NC-4.0).

@misc{itaim2026seeing,
  title  = {Seeing to Drive on a Single Board: Attention-Free Distilled Depth on a Raspberry Pi 5 + Hailo-8L
            for Onboard Obstacle Avoidance, Planning and SLAM},
  author = {M., Itai},
  year   = {2026},
  url    = {https://itaim18.github.io/seeing-to-drive/}
}