README.md

---
license: cc-by-4.0
task_categories:
- robotics
tags:
- autonomous-driving
- nurec
- closed-loop-eval
- alpasim
- alpamayo
- mads
size_categories:
- n<1K
---

# Nurec_eval_v2 — 59 OOD-longtail scenes (MADS schema, trajdata-compatible)

**59 NRE 26.02 closed-loop eval scenes** with HD-map data packaged in the **MADS schema** that [trajdata](https://github.com/NVlabs/trajdata)'s `populate_vector_map()` reads. Drop-in usable by [NVlabs/alpasim](https://github.com/NVlabs/alpasim) — when you call `Artifact(usdz).map`, you get back a fully-populated `VectorMap` (lanes / road edges / wait lines) and the driver model gets **lane-aware** input.

## What changed vs v1 (Nurec_eval)

| | v1 (Nurec_eval) | v2 (this) |
|---|---|---|
| Map artifact location | `map_data/cf_*.parquet`, `dw_lane.parquet`, ... | `map_data/{lane,road_boundary,association,wait_line,traffic_sign,clip}.parquet` |
| Schema | NRE-internal **ClipGT** format | **MADS** format ([trajdata expects this](https://github.com/NVlabs/trajdata/blob/main/src/trajdata/dataset_specific/mads/mads_utils.py#L256)) |
| alpasim `Artifact.map` loads it | ❌ KeyError on `lane.parquet` | ✅ Loads ~30–280 lanes / 10–30 road edges per scene |
| Driver model has map context | ❌ map = None | ✅ VectorMap populated |

If you only need raw NRE ClipGT data, use [luuuulinnnn/Nurec_eval](https://huggingface.co/datasets/luuuulinnnn/Nurec_eval) (v1).

## What's inside each `pai_<uuid>.usdz`

| File | Purpose |
|---|---|
| `default.usda`, `parsed_config.yaml`, `data_info.json`, `metadata.yaml`, `datasource_summary.json` | Scene metadata |
| `checkpoint.ckpt` | NRE-26.02 Gaussian-splat reconstruction weights (~700–800 MB) |
| `mesh.ply`, `mesh.usd` | Poisson reconstruction mesh |
| `ground_mesh.ply` | Ground mesh (for collision_with_ground / offroad-by-ground) |
| `rig_trajectories.json`, `rig_trajectories.usda` | Ego recorded trajectory (used by `route_generator_type=RECORDED`) |
| `sequence_tracks.json`, `sequence_tracks.usda` | Other actors' recorded trajectories (replay mode for trafficsim) |
| **`map_data/clip.parquet`** | **MADS: scene metadata** |
| **`map_data/lane.parquet`** | **MADS: lanes (key.map_id, Lane.left_rail, Lane.right_rail)** |
| **`map_data/road_boundary.parquet`** | **MADS: road edges (RoadBoundary.location polyline)** |
| **`map_data/association.parquet`** | **MADS: NEXT_LANE / PREVIOUS_LANE / LEFT_LANE / RIGHT_LANE relations** |
| **`map_data/wait_line.parquet`** | **MADS: wait lines (from NRE crosswalks)** |
| **`map_data/traffic_sign.parquet`** | **MADS: traffic signs (synthetic; NRE recon doesn't ship signs)** |

## Scene composition — 59 scenes from 8 OOD-longtail buckets

| Category | Scenes |
|---|---:|
| Animals / Birds / Roadkill | 7 |
| Complex Intersection Interaction | 8 |
| Cyclists & Micromobility Complex | 8 |
| Emergency Incident Scene | 7 |
| Pedestrian Density / Close Proximity | 8 |
| Road Debris / Safety Traces | 5 |
| Special / Uncommon Vehicle Behavior | 8 |
| Work Zones / Temp Traffic Control | 8 |
| ⚠️ Broken route (waypoint folds back) | 1 |
| **Total usable** | **58 scenes** |

Scene → category mapping is in `category_uui.json`.

## Usage with AlpaSim

```bash
alpasim_wizard \
  deploy=local topology=1gpu driver=<your_model> \
  +physics=disabled +vehicle=custom \
  scenes.path=/path/to/Nurec_eval_v2 \
  scenes.scenes_csv=/path/to/Nurec_eval_v2/sim_scenes.csv \
  scenes.suites_csv=/path/to/Nurec_eval_v2/sim_suites.csv \
  ...
```

Map will be auto-loaded by `Artifact.map` and fed to the driver as part of scenario context.

## How the MADS files were produced

The included [`convert_cf_to_mads.py`](./convert_cf_to_mads.py) converts NRE ClipGT parquets
(`cf_crosswalks.parquet`, `cf_lane_topology_node.parquet`, `cf_road_boundary.parquet`,
`dw_lane.parquet`, `lane_chunk.parquet`, `lane_rail.parquet`) into the 6 MADS files.

Then [`inject_mads_map.py`](./inject_mads_map.py) zips them into the USDZ at `map_data/`.

Key mappings:
- `dw_lane.currentId + chunkIndices` + `lane_chunk.left/right` → `lane.parquet`
  (`Lane.left_rail`, `Lane.right_rail`)
- `cf_road_boundary.road_boundary_polyline` → `road_boundary.parquet`
  (`RoadBoundary.location`)
- `dw_lane.lgwm_lane_continuation_array` → `association.parquet`
  (`NEXT_LANE` / `PREVIOUS_LANE`)
- `lane_rail.lane_indices` → `association.parquet`
  (`LEFT_LANE` / `RIGHT_LANE`)
- `cf_crosswalks.reference_line` → `wait_line.parquet`
  (`WaitLine.location`, `WaitLine.category = "CROSSWALK"`)

## Source

- Scenes exported from NRE 26.02 (`26.2.158-2b8da993`)
- HD map from NRE `ood_closeloop_map`, converted CF → MADS by `convert_cf_to_mads.py`
- Compatible with [NVlabs/alpasim](https://github.com/NVlabs/alpasim) (commit reading `map_data/lane.parquet`)
- Companion datasets: [luuuulinnnn/OOD17_longtail_usdz](https://huggingface.co/datasets/luuuulinnnn/OOD17_longtail_usdz), [luuuulinnnn/OOD64-NuRec-USDZ-with-ground](https://huggingface.co/datasets/luuuulinnnn/OOD64-NuRec-USDZ-with-ground), [luuuulinnnn/Nurec_eval](https://huggingface.co/datasets/luuuulinnnn/Nurec_eval) (CF-format v1)