# Sample inputs shipped with diffusion-planner-p150

Only REDISTRIBUTABLE data goes here (Apache-2.0 / MIT / CC-BY with attribution). Data whose license is unstated or
non-commercial (the Autoware demo rosbag, nuScenes, Argoverse 2, ...) is NOT shipped; the nuScenes-derived planner
instants used for local agreement tests live in the porting workspace only (`research/diffusion-planner/public_data`).

Never use the `.bin` suffix here (or `.pt`, `.pth`, `.ckpt`, `.safetensors`): tt-model's staging silently drops
those suffixes from `code/` (`CODE_IGNORE`, tt-model-manager `src/tt_kernel/build.py:328-332`).

Each sample is one `.npz` holding exactly the 15 raw (pre-normalization, ego-frame, batch-1, float32) tensors of the
node's `DiffusionPlannerCore::create_input_data()` (`reference/config.py` `INPUT_SCHEMA`): pass it as
`model(inputs="<file>.npz")` or as the `/predict` field `inputs` (`server/client.py --inputs <file>.npz`).

Next to each sample, `<stem>.reference.json` is the `/predict` body of the fp32 CPU reference
(`tt_diffusion_planner.reference.ReferencePlanner`, `Output.to_dict()`, timing removed) on it;
`server/smoke_test.py` compares the served trajectory with it (ADE / FDE gates). Regenerate both the reference bodies
and `tests/goldens/` with `code/scripts/ref_golden.py` whenever the reference, the weights or the post-processing
changes.

| file | content | source | license |
|---|---|---|---|
| `kashiwanoha_dense.npz` (default sample) | one planning instant on the kashiwanoha test map: ego at 6 m/s on a 17-lanelet route, 88 neighbours (constant-speed vehicles), 123 lanes (24 lanelets with traffic lights in the map), 60 line strings (stop lines and road borders), no intersection polygon; 113,893 B, sha256 `d8c2aaef...99f7` | Lanelet2 map [AutowareFoundation/map-carla-kashiwanoha](https://huggingface.co/datasets/AutowareFoundation/map-carla-kashiwanoha) 0.2.0 (`lanelet2_map.osm`, sha256 `4fe358f2...4e6d`, byte-identical to autoware_universe `planning/autoware_diffusion_planner/test_map`); scene scripted by `research/diffusion-planner/scripts/dp_scene.py` (`dp_reference.py --scene osm --osm .../lanelet2_map.osm --n-agents 120 --seed 7`), a Python port of the node's tensor construction (simplifications: SPEC 7) | Apache-2.0 (map: Apache-2.0 per its dataset card) |
| `straight_road.npz` | procedural 3-lane straight road: ego at 8 m/s, 12 neighbours (vehicles and pedestrians), 33 lanes with a green traffic light ahead, 6 route lanes, 1 intersection polygon, 9 line strings (borders + stop line); the precision-sensitive scene of SPEC 0.5; 10,832 B, sha256 `b9979b90...d381` | generated by `research/diffusion-planner/scripts/dp_scene.py` (`dp_reference.py --scene straight`), no external data | Apache-2.0 (this repo) |