OCCWAM B2D v9 step18000

This repository contains the deployed OCCWAM joint flow-matching checkpoint used by the v9 Bench2Drive closed-loop agent.

Checkpoint

Field	Value
File	v9_step18000/v9step18000_FULL_merged.pt
Size	1,656,918,530 bytes
SHA-256	43d914c7afa8c892b03023c96a6f3619b1aa37d874e103d28a12824a1c42316f
Envelope	occwam-fullmot-best-model-v1
Global step	18000
Config SHA-256	7948cd560c83dc2b519be5ff882c5ff68f43daf3456e4fb774d4ae659017651b

Training and inference code: https://github.com/HQH111/OCCWAM

The inference graph is scene-frame joint OCC/trajectory flow matching from white noise, followed by a hard scene-to-future-ego warp, VAE re-encoding and one velocity-parameterised ego-rollout refinement.

Required contract

export AEC_TRAJ1_CTX=2
export AEC_TRAJ2_VFM=1
export OCCWAM_TP_NORM_UNIFIED=1

The unified target-point normalisation flag must be set at both training and inference. Omitting it silently changes the input conditioning.

This checkpoint does not embed the frozen Qwen3-VL, Stage-3 carrier, OCC VAE, TransFuser++ and Profile-B assets. Their hashes and roles are documented in the GitHub repository's runtime and training configs.