2026 ETRI Autonomous Driving Challenge — Checkpoints
VAD/LAW checkpoints for the 2026 ETRI E2E Driving Challenge. Full training/eval
code lives on GitHub at gyuz1/2026_Etri_challenge
(branches LAW-split, LAW-fulldata) — each checkpoint below was produced by
the config file of the same name in this repo, which is a snapshot of the
exact config used for that run.
Checkpoints
| File | Produced by | Source |
|---|---|---|
law_pretrained_nus.pth |
config_law_pretrained_nus.py |
Perception-based LAW, pretrained on nuScenes (not yet pushed to GitHub — local project LAW_p-based) |
stage1_etri_fulldata_10hz_epoch48.pth |
config_stage1_etri_fulldata_10hz.py |
LAW-fulldata branch, fulldata/configs/VAD_etri_tiny_stage1_cached_fulldata_10hz.py |
stage1_etri_split_301_75_2hz_epoch48.pth |
config_stage1_etri_split_301_75_2hz.py |
LAW-split branch, projects/configs/VAD/VAD_etri_tiny_stage1_cached.py |
Naming convention for every checkpoint here: stage{N}_{split_XXX_YY|fulldata}_{N}hz_epoch{E}.pth — stage, data scope, ego-motion Hz, and epoch are always explicit.
law_pretrained_nus.pth
LAW world-model pretraining on nuScenes (perception-based variant: joint
detection + map + waypoint loss, L_agent + L_map + loss_rec +
loss_waypoint). Used as the ImageNet/ResNet50-equivalent initialization
(load_from) for all ETRI VAD/LAW stage1 runs below.
stage1_etri_fulldata_10hz_epoch48.pth
VAD-tiny stage1 (detection + agent motion + map only, planning loss = 0),
trained on the full 376-scene ETRI dataset with 10Hz ego-motion annotations
(dense causal robust-polynomial-fit velocity/acceleration, per the
organizers' 2026-08-25 Q&A). 48/48 epochs, load_from=law_pretrained_nus.pth.
stage1_etri_split_301_75_2hz_epoch48.pth
VAD-tiny stage1 (detection + agent motion + map only, planning loss = 0),
trained on the 301/75 scene held-out split with 2Hz ego-motion annotations
(the original causal-regen split, predating the 10Hz/teammate-split
refinements). 48/48 epochs, load_from=law_pretrained_nus.pth. This is the
stage1 checkpoint stage2 runs initialize from.
Stage2 (VADLAW) checkpoints are not currently published here -- the target_point-attention-conditioned architecture they were trained with is being replaced (organizer ruling: target_point may only be used to select among already-generated candidates, never fed into the network) and the 301/75-split checkpoint trained under that architecture was removed.