---
library_name: pytorch
tags:
- autonomous-driving
- flow-matching
- occupancy-prediction
- trajectory-prediction
- bench2drive
---

# 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

```bash
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.