JEPA-Shield runtime assets

Model weights and Gaussian renderer assets for JEPA-Shield, maintained by Taekyung Kim.

This release contains the Python car, GS car, Python tractor-trailer, and GS tractor-trailer deployments. Each deployment includes the frozen Le-WM, learned occupancy and motion models, preprocessing, controller configuration, and parking scenario bank. Tractor-trailer bundles include their runtime profiles. The Gaussian folder contains the scene, vehicle and trailer PLYs and dimension metadata.

Safety-Probe weights are available for Python car, GS car, and Python tractor-trailer. GS tractor-trailer Safety-Probe is not implemented. GS tractor-trailer is supported as an experimental task and is excluded from the paper comparison.

Download

After installing the source repository, run from its root:

uv run python scripts/download_assets.py

The downloader installs the pinned release into assets/ after verifying every file's size and SHA-256 against manifest.json.

Image-goal evaluation

calibration/latent_success.yaml contains the original calibrated thresholds for Python car, GS car, and Python tractor-trailer, bound to the exported encoder checkpoint identities. The thresholds have not been recalibrated. GS tractor-trailer requires a user-supplied threshold or calibrated configuration; no GS tractor-trailer image-goal results are claimed.

Image-goal case construction requires a locally generated held-out driving dataset. Dataset generation and training instructions are in the source repository. Datasets are not distributed here.

Export format

Checkpoints contain runtime tensors, model schemas, and normalization statistics. Every tensor was compared with its source checkpoint during export. Optimizer state, training logs, and private training metadata are omitted. Deployment manifests retain source checkpoint hashes and identify the derived runtime files without claiming a new benchmark evaluation. The release contains 86 runtime files, approximately 2.60 GB.