ZhidongWang/seerdrive-pdm-lite
0
SeerDrive (pdm_lite / Bench2Drive) -- D-23 GT-nearest
End-to-end driving model (NeurIPS-2025 SeerDrive) ported to CARLA pdm_lite / Bench2Drive (geometric-path + speed-bin, BridgeDrive-style).
Checkpoint
seerdrive_ep50.pt-- model weights, epoch 50.
Regime (D-23)
- 256x256 ego-centered surround BEV (rear/side included).
- Rear/side cut-in threat reward folded into NC/TTC.
- GT-nearest speed-only simreward (`simtrainimshape=False`): geometry is imitation; reward supervises the speed head at the GT-nearest expert anchor.
- Frozen backbone: carlagarage TransFuser (regnety032, alltowns, `model0030_2.pth`).
Training
- 8x A100 DDP, 50 epochs, effective batch 128, cosine LR 1e-4 -> ~1e-6.
