HeavyBridge checkpoints

Checkpoints for

Sotirios Chatzis and Loukas Papadoulas. HeavyBridge: Tempered Lévy Bridges for Anchored Diffusion Driving Policies. Transactions on Machine Learning Research, 2026. OpenReview

Code, per-scene results and documentation: github.com/soteri0s/heavybridge. The variant ids below are explained in its README.

checkpoints/

DiffusionDrive planners with the HeavyBridge trajectory head, warm-started from the public DiffusionDrive checkpoint and trained on navtrain. Each file is a PyTorch Lightning checkpoint ({"state_dict": {"agent.*": ...}}), loadable by DiffusionDrive's TransfuserAgent with apply_patch.py active and HB_MIXING (plus HB_ALPHA, HB_KAPPA, HB_NU where relevant) set as in the variant table of train_stage.sh.

file	variant	recipe
R3gfull_final.pth, R3gfull_s2_final.pth	R3g-full, the proposed planner (per-step tempered stable, α = 1.8, κ = 0.956338), seeds 1 and 2	100 epochs
R0full_final.pth, R0full_s2_final.pth	R0-full, Gaussian bridge, seeds 1 and 2	100 epochs
R0_finetune_ep15.pth	R0, Gaussian bridge	15-epoch screening
R1a_final.pth, R1b_final.pth	per-rollout Inverse-Gamma, ν = 5 / 10	15-epoch screening
R2_final.pth	per-rollout scene-amortized scale (R2c0 is this checkpoint with the churn off)	15-epoch screening
R2b_final.pth	bounded amortized per-rollout scale (control)	15-epoch screening
R3t_final.pth, R3g_final.pth, R3k_final.pth, R23_final.pth	per-step variants of the screening phase	15-epoch screening
R3g2_final.pth, R3gk_final.pth, R23g_final.pth	R3g′, R3gk, R23′ of the navtest sweep	15-epoch screening
R3ga7_final.pth, R3ga6_final.pth	α = 1.7 / 1.6 at jump energy 5/4 (α ablation)	15-epoch screening

The raw selection files and training logs are available from the authors on request.

The checkpoints are trained on, and the outputs derived from, NAVSIM (OpenScene / nuPlan) data, and are subject to those datasets' licenses. HeavyBridge builds on DiffusionDrive (MIT).

Citation

@article{chatzis2026heavybridge,
  title   = {HeavyBridge: Tempered L{\'e}vy Bridges for Anchored Diffusion Driving Policies},
  author  = {Chatzis, Sotirios and Papadoulas, Loukas},
  journal = {Transactions on Machine Learning Research},
  year    = {2026},
  url     = {https://openreview.net/forum?id=5p1S532fXM}
}

This work was funded by the European Union under the HIDDEN project, grant agreement No 101202228.