ActWorld-JEPA
Inference weights and frozen learned selectors for ActWorld-JEPA: Action-Conditioned JEPA World Models Meet Latent Planning for End-to-End Autonomous Driving.
Code and instructions · Project page · Evaluation protocols
| Deployment | Artifacts | Reported main result |
|---|---|---|
| NAVSIM v1 R383 | actworld_jepa_navsim_v1_r383_bundle.tar.gz (checkpoint plus four selectors) |
Author-reported leaderboard PDMS 90.07 |
| NAVSIM v2 NAVTEST | actworld_navsim_v2.ckpt, factor005.pkl, strict_variant_selector.pkl, multicomponent_gate_selector.pkl |
Corrected EPDMS 91.5900 |
| NAVSIM v2 NAVHARD | NAVTEST artifacts plus temporal_fcts_pair_head.pkl, temporal_fcts_head.pt; deployment scripts in GitHub |
Corrected combined EPDMS 42.7351 |
| nuScenes round 13 | actworld_nuscenes_r13.pth |
Repeat 04: 6 s L2 2.3945 m; six-column mean 1.204533 m / 0.763833% box collision |
Download
hf download shengkaizhang/ActWorld-JEPA --local-dir weights
Check SHA256 and sizes against manifest.json. Extract the v1 bundle for its
checkpoint and selector files. Follow benchmark-specific GitHub instructions;
these are custom research models, not Transformers from_pretrained models.
The v2 and nuScenes checkpoints are state-dict-only exports. All 587 / 1,414 model tensors respectively were checked equal to the evaluated checkpoints. Optimizer states, training callbacks and private path metadata are excluded. Selectors retain learned estimators/reference features required for inference. Only load pickle artifacts from sources you trust.
Evaluation and limitations
NAVSIM v2 corrected scores recompute human-filter aggregate terms. The same predictions previously scored 88.1771 on NAVTEST and 39.2045 on NAVHARD. This is an evaluator correction, not a model improvement. Compare with other methods only under matched evaluator versions, splits and aggregation.
R383's earlier public release records local PDMS 94.0721. The authors identify that submission as leaderboard 90.07/local 90.05; the distinct records are preserved in the protocol documentation rather than silently relabelled. nuScenes repeat 04 was selected from repeated inference; last-decimal values can vary with execution. Checkpoint-only evaluation omitting the deployed selectors/postprocessing does not reproduce the full model.
These models are intended for research. They are not validated for deployment on public roads. Dataset providers retain their data rights; no benchmark sensor data or metric caches are included. ActWorld code/artifacts are released under Apache-2.0, with inherited dependencies retaining their own licenses. The source repository includes third-party notices for NAVSIM, Drive-JEPA, V-JEPA 2, MomAD and SparseDrive.