ZhidongWang/seerdrive-b2d

SeerDrive — Bench2Drive / pdm_lite port

End-to-end driving model: SeerDrive (NeurIPS 2025) ported onto CARLA pdm_lite (carla_garage / Bench2Drive). Geometric-path + speed-bin formulation (BridgeDrive-style).

Pipeline (waypoint -> destination)

TransFuser BEV features [1512,8,8] + 12-dim ego status (incl. navigation target_point) -> SeerDrive (21M params) predicts a 20-point geometric path (lateral) + a 9-bin target speed (longitudinal, RL-tuned) -> controller -> throttle/steer/brake -> closed-loop to goal.

• Path branch: 60 k-means shape anchors + offset decoder (imitation-selected)
• Speed branch: 9-bin head; deploy argmax(speed_logit + lambda*reward)
• World model: predicts future BEV at path completion (self-supervised)
• RL on speed: NC/DAC/EP/TTC/Comfort box-geometry sim-reward heads

Training

30 epochs, per-GPU bs16 x 8 GPUs (eff_bs 128), full pdm_lite (~475k train / 120k val frames), cosine LR 1e-4, Adam.

Final val metrics (ep30)

geo_ade	geo_fde	speed_acc	speed_l1	mIoU	mIoU_fut
0.058 m	0.113 m	0.946	0.547	0.416	0.308

IoU: road 0.55 / lane 0.37 / veh 0.28 / ego 0.97.

Files

• seerdrive_ep30.pt — final model weights + val metrics (~85 MB)
• seerdrive_last.pt — full state (model + optimizer + scheduler) for --resume (~254 MB)