TinyJEPA v1: Simulation Data, Firmware, Training Runs and Logs for a JEPA World Model that Plans and Learns on an ESP32-S3
Description
Data, source code, ESP32-S3 firmware, training runs and logs accompanying the preprint "A World Model in 41k Parameters: Planning and On-Chip Learning with a JEPA on an ESP32-S3".
The record contains: three HDF5 datasets collected in the Webots simulator with a mecanum robot (96x96 grayscale camera, four sonars, IMU): 600 training episodes (82,442 transitions), 100 validation episodes (14,050) and 100 confirmation episodes (14,533), on disjoint map seeds; the Webots world, data collector, closed-loop planner and training script; the Colab notebook that ran every training, with the hypotheses written before each run, and the reports and checkpoints of all 20 runs; the ESP-IDF firmware of every pre-specified gate (P0 feasibility, P4a INT8 parity, P4b chip-in-the-loop, P5 on-chip adaptation); the deployed model checkpoint (SHA-256 b597a7ed87dffc210e04168edee456106926a7d68dab965ea1dedc21272790dc, printed by the firmware at boot); the CSV logs of every closed-loop experiment, device console logs, the analysis scripts fixed before each run and their outputs; and a script that regenerates every data figure and computed number of the paper.
Failed and inconclusive gates are reported and their data are included. SHA-256 checksums are in SHA256SUMS.txt inside the archive and, for the HDF5 files, in its README.
Licences: code under MIT; data, logs, figures, model and documentation under CC BY 4.0.