| --- |
| license: cc-by-nc-4.0 |
| task_categories: |
| - text-generation |
| pretty_name: AutoMoT PDM-Lite BEV Encoder Indexes |
| --- |
| |
| # AutoMoT PDM-Lite BEV Encoder Indexes |
|
|
| This dataset provides the prepared PDM-Lite JSONL indexes for AutoMoT training. |
|
|
| ## Files |
|
|
| - `pdm_lite_2hz_2tp_train_bev_encoder.jsonl` |
| - `pdm_lite_2hz_2tp_val_bev_encoder.jsonl` |
|
|
| Each row contains four historical front-camera paths in `image`, the current |
| front-camera path in `front`, trajectory and route supervision, future-speed |
| supervision, and a reference to the precomputed current-frame BEV feature: |
|
|
| - `bev_encoder_feature` |
| - `bev_encoder_feature_frame` |
|
|
| The prompt contains four `<image>` tokens for Qwen3-VL reasoning and one |
| `<bev>` token for the action branch. The BEV feature path is relative to |
| `PDM_DATA_DIR` and points to: |
|
|
| ```text |
| <PDM_DATA_DIR>/<scenario>/<route>/bev_encoder_feature/route_features.pt |
| ``` |
|
|
| AutoMoT reads 64 spatial BEV tokens (`8 x 8`) from this file. If the cache is |
| unavailable, `front` identifies the RGB frame used by the online BEV encoder |
| fallback together with the corresponding LiDAR BEV input. |
|
|
| Training instructions are available in the |
| [AutoMoT repository](https://github.com/OscarHuangWind/AutoMoT/tree/release-training). |
|
|
