melloyello/asl-trunk-world-model

ASL soft trunk robot: world-model dataset and training package

Research data of the Stanford ASL trunk robot project (recorded 2026-09-10/11). Code: https://github.com/mellowyellow71/trunk-stack (branch world-model, a fork of StanfordASL/trunk-stack). Everything a new person needs to reproduce or continue the work is in this repository plus that branch; no access to the original machines is required.

What is here

path	content
train_package_2026-09-11.tar.gz (+ .sha256)	self-contained training package: DINO-WM code (commit 0a9492f + trunk loader/config/patch), the dataset, DINOv2 and decoder weights, launch.sh, preflight.sh, SLURM template, report template, handoff PDF. Start here.
handoff_training_b300.pdf	the operator handoff: recipe, node requirements, procedure, decision rules, failure modes, delivery
dataset/trunk/{train,val}	the same dataset unpacked (DINO-WM layout): 224x224 MP4 at 50 Hz, abs_actions.pth [N,T,4], states.pth [N,T,12], velocities.pth, timestamps.pth, seq_lengths.pkl, index.json
bags/	the raw rosbag2 recordings (mcap): Phase A 1001-1003 and Phase B 2101, 2001-2012, 26 GB; topics /image_raw/compressed, /all_motors_control, /all_motors_status_stamped, /motor_currents, /motor_temperatures; rebuild the dataset with build_dinowm_dataset.py from the code repo
docs/	session notes (recording, checks, sanity model), camera calibration (2.446 mm/px at 224 px), per-file check reports, the training plan and the project handoff

Dataset: 13 recordings of 304 s each, 4 active motors (commands in degrees relative to rest, +-90), camera-only observations. Train: 45 episodes / 167,417 frames (55.8 min). Val: 6 episodes / 30,493 frames (a held-out mixture file and the step-only file). Frames resampled to a uniform 50 Hz grid; motor state interpolated between bracketing 100 Hz reads (gap <= 40 ms); commands are held values.

Use

hf download melloyello/asl-trunk-world-model --repo-type dataset --local-dir asl-trunk-world-model
cd asl-trunk-world-model && sha256sum -c train_package_2026-09-11.tar.gz.sha256
tar xzf train_package_2026-09-11.tar.gz && cd train_package_2026-09-11 && ./preflight.sh

Then follow handoff_training_b300.pdf. Results go back under runs/ in this repository.