well-iam/humanoid-intern-challenge

Pushing world model: my clips and cached results

Data and cached results for the pushing-world-model project. A small world model predicts, in DINOv2 feature space, where an object goes when a pusher moves. It is trained on Push-T, then on 115 clips of my own hand pushing a box on a table. My recorded palm paths then drive a simulated Franka Panda in MuJoCo.

The project's Colab notebook downloads this dataset and rebuilds every figure and GIF from it, without a GPU.

from huggingface_hub import snapshot_download

snapshot_download("well-iam/humanoid-intern-challenge", repo_type="dataset", local_dir="pwm-data")

What is in it

My clips (clips/), recorded on 16 October 2026 with a phone 71 cm above a dark carpet, pushing a white box (17.8 x 5.8 x 3 cm) inside a 50 x 50 cm square. 9 takes, 119 clips cut at claps; 4 junk clips left out (the hand resetting the box; the list and reasons are in the tracking report), so 115 clips: 80 pushes, 35 near-misses.

path	what
clips/videos/<take>.mp4	the 9 takes, 720p H.264 with audio (the claps cut the takes into clips)
clips/episodes/<clip>.npz	one episode per clip: the straightened 512 x 512 px top-down frames at 8 fps, palm and box tracks (cm), what was seen, the box angle, the hand direction, the split and the kind (push or near-miss)
clips/tracks.csv	the same tracks as one table, one row per frame
clips/split.json	the held-out split: 18 clips (10 pushes, 8 near-misses): the first 3 takes' 8, plus 4 pushes and 6 near-misses of the later takes drawn with a fixed seed
clips/tracking/	the tracking report and one overlay GIF per take
clips/dinov2/embeddings/<clip>.npy	DINOv2-small patch embeddings, fp16, (frames, 256, 384)
clips/dinov2/predictor.pt, predictor-blind.pt	the trained predictor and the action-blind predictor
clips/dinov2/	results, reports and figures: error vs horizon, near-miss test, retrievals, training curves, probe error vs number of training clips (probe-scaling.png)
clips/dinov2/only-takes/	the same model scored on the first 3 takes' 8 held-out clips only (for the before/after)
clips/panda/	the Panda replay of every clip (replay.json) and GIFs
clips/dinov2-30clips/, clips/panda-30clips/	the first run, on the first 30 clips (3 takes): results, figures, predictors, replay
clips/gif-strip/	a held-out clip, the predictor's view one second ahead (nearest real frame) and the Panda replay, side by side
clips/logs/	training logs (batch loss every 100 steps)
clips/aruco-a4.pdf	the printed marker that fixes the table frame

Table frame: cm, origin at the square's top-left corner as seen from the camera, x to the right, y down. Palm = mean of MediaPipe hand landmarks 0, 5, 9, 13, 17, height-corrected.

Push-T results (pusht/): DINOv2-small embeddings of the 150 Push-T episodes used (fp16, pusht/dinov2/embeddings/pusht-<index>.npy, index = position in Push-T's file), the trained predictors (three seeds for DINOv2), results, the ranking test and the Panda replay. For V-JEPA 2 ViT-L, the encoder that lost, only the predictors, results and figures: its 9.3 GB of embeddings are left out (they are only needed to retrain it; scripts/embed.py --dataset pusht --encoder vjepa2 recomputes them). Push-T itself is not included: the code downloads it from its original URL (https://diffusion-policy.cs.columbia.edu/data/training/pusht.zip, from Chi et al., Diffusion Policy, 2023).

Code, the exact commands and the results are in the project's repository and README.

License

CC BY 4.0. The Push-T embeddings are computed from the Push-T dataset; its own terms apply to the underlying data.