File size: 2,417 Bytes
e790676
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
20f215a
 
 
e790676
20f215a
 
 
 
e790676
 
 
c9f4be0
20f215a
e790676
 
 
 
 
 
 
c9f4be0
e790676
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
---
license: cc-by-nc-sa-4.0
pretty_name: ReactBench-AD
tags:
  - autonomous-driving
  - end-to-end-driving
  - vision-language-action
  - benchmark
  - nuplan
  - navsim
---

# ReactBench-AD

[Paper](https://arxiv.org/abs/2611.20011) | [Code](https://github.com/niktheod/reactbench-ad)

<p align="center">
  <img src="https://huggingface.co/datasets/niktheod/reactbench-ad/resolve/main/docs/reactbench_overview.png" alt="ReactBench-AD overview" width="100%">
</p>

ReactBench-AD composites 3D assets into real NAVSIM (**ReactNAV**, 3,228 navtest samples) and
PhysicalAI-Autonomous-Vehicles (**ReactPAI**, 1,045 clips) scenes along the ego vehicle's logged trajectory and
measures the reaction of driving models by comparing the trajectory they predict for the original scene
(y<sup>O</sup>) with the one for the edited scene (y<sup>E</sup>), isolating the causal effect of the inserted object.

## Files

- `reactnav/<object>_tau<τ>s.tar.gz`: the 12 edited variants, a `crossing_pedestrian`, `standing_pedestrian` or `car` at
  look-ahead time τ = 0.5, 1.0, 1.5 or 2.0 s. Each holds the camera frames the inserted object changes and
  one metadata file per sample describing the object. The unchanged frames are the original navtest images.
- `reactnav/lanes.tar`: the NAVSIM lane polygons the ReactNAV scorer uses.

## Usage

ReactNAV builds on your copy of NAVSIM navtest. Follow
[Getting the data](https://github.com/niktheod/reactbench-ad#getting-the-data) in the code repository: extract the
variant archives into `datasets/reactnav/` and `lanes.tar` into `outputs/work/navsim/`, then
`python -m reactbench.datasets.reactnav link` adds the unchanged frames from navtest.

```bash
hf download niktheod/reactbench-ad --repo-type dataset --local-dir reactbench-ad
```

## License

CC BY-NC-SA 4.0. ReactNAV is derived from nuPlan, OpenScene and NAVSIM and is released under the license of those
datasets.

## Citation

```bibtex
@misc{theodoridis2026reactbenchaddrivingvlasworldaction,
      title={ReactBench-AD: Do Driving VLAs and World-Action Models React to Imminent Hazards in Their Path?},
      author={Nikos Theodoridis and Reenu Mohandas and Sushil Sharma and Ciarán Hogan and Ganesh Sistu and Anthony Scanlan and Tim Brophy and Ciarán Eising},
      year={2026},
      eprint={2611.20011},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2611.20011},
}
```