Update model card for Pseudo-Simulation (NAVSIM v2)
#3
by
nielsr
HF Staff
- opened
README.md
CHANGED
|
@@ -3,14 +3,45 @@ license: apache-2.0
|
|
| 3 |
pipeline_tag: robotics
|
| 4 |
---
|
| 5 |
|
| 6 |
-
|
| 7 |
|
| 8 |
-
<
|
| 9 |
<img src="https://raw.githubusercontent.com/autonomousvision/navsim/main/assets/navsim_transparent.png" width="400">
|
| 10 |
-
<
|
| 11 |
-
|
| 12 |
-
</
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 13 |
|
| 14 |
-
|
| 15 |
|
| 16 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 3 |
pipeline_tag: robotics
|
| 4 |
---
|
| 5 |
|
| 6 |
+
# Pseudo-Simulation for Autonomous Driving
|
| 7 |
|
| 8 |
+
<div id="top" align="center">
|
| 9 |
<img src="https://raw.githubusercontent.com/autonomousvision/navsim/main/assets/navsim_transparent.png" width="400">
|
| 10 |
+
<h3 align="center">
|
| 11 |
+
<a href="https://huggingface.co/papers/2506.04218">Paper</a> |
|
| 12 |
+
<a href="https://vveicao.github.io/projects/NavsimV2/Cao2025_supp.pdf">Project Page</a> |
|
| 13 |
+
<a href="https://github.com/autonomousvision/navsim">GitHub</a> |
|
| 14 |
+
<a href="https://huggingface.co/spaces/AGC2025/e2e-driving-navhard">Public Leaderboard v2 (navhard)</a> |
|
| 15 |
+
<a href="https://huggingface.co/spaces/AGC2025/e2e-driving-2025">Challenge Leaderboard v2 (ICCV)</a>
|
| 16 |
+
</h3>
|
| 17 |
+
</div>
|
| 18 |
+
|
| 19 |
+
This repository provides official model checkpoints and code for **Pseudo-Simulation**, a novel paradigm for evaluating Autonomous Vehicles (AVs), as introduced in the paper "[Pseudo-Simulation for Autonomous Driving](https://huggingface.co/papers/2506.04218)". This work was presented at CoRL 2025.
|
| 20 |
+
|
| 21 |
+
Pseudo-simulation addresses critical limitations in existing AV evaluation paradigms. It combines the efficiency of open-loop evaluation with the robustness of closed-loop evaluation by operating on real datasets and augmenting them with synthetic observations generated prior to evaluation using 3D Gaussian Splatting. The key idea is to approximate potential future states the AV might encounter by generating a diverse set of observations that vary in position, heading, and speed. A novel proximity-based weighting scheme then assigns higher importance to synthetic observations that best match the AV's likely behavior. This enables evaluating error recovery and the mitigation of causal confusion, similar to closed-loop benchmarks, but without requiring sequential interactive simulation.
|
| 22 |
+
|
| 23 |
+
Pseudo-simulation achieves a strong correlation ($R^2=0.8$) with traditional, computationally expensive closed-loop simulations while being much faster and easier to scale (requiring 6x less compute). Unlike traditional closed-loop simulation, pseudo-simulation is neither sequential nor interactive, enabling the open-loop computation of all evaluation metrics in the associated leaderboards. It serves as the primary evaluation framework for the [AGC2025 NAVSIM End-to-End Driving Challenge](https://huggingface.co/spaces/AGC2025/e2e-driving-2025).
|
| 24 |
+
|
| 25 |
+
This repository contains official model checkpoints for autonomous driving agents such as TransFuser, Latent TransFuser (LTF), and the EgoStatusMLP. These checkpoints were used to populate the [leaderboards](https://huggingface.co/spaces/AGC2025/e2e-driving-navhard) and demonstrate the Pseudo-Simulation framework's capabilities.
|
| 26 |
|
| 27 |
+
For further information including installation, understanding and creating agents, data formats, metrics, and submission guidelines, please visit the official [NAVSIM GitHub repository](https://github.com/autonomousvision/navsim).
|
| 28 |
|
| 29 |
+
## Citation
|
| 30 |
+
If you find our work helpful or inspiring, please feel free to cite our latest paper:
|
| 31 |
+
```bibtex
|
| 32 |
+
@inproceedings{Cao2025CORL,
|
| 33 |
+
author = {Wei Cao and Marcel Hallgarten and Tianyu Li and Daniel Dauner and Xunjiang Gu and Caojun Wang and Yakov Miron and Marco Aiello and Hongyang Li and Igor Gilitschenski and Boris Ivanovic and Marco Pavone and Andreas Geiger and Kashyap Chitta},
|
| 34 |
+
title = {Pseudo-Simulation for Autonomous Driving},
|
| 35 |
+
booktitle = {Conference on Robot Learning (CoRL)},
|
| 36 |
+
year = {2025},
|
| 37 |
+
}
|
| 38 |
+
```
|
| 39 |
+
You may also find the related NAVSIM v1 paper useful:
|
| 40 |
+
```bibtex
|
| 41 |
+
@inproceedings{Dauner2024NEURIPS,
|
| 42 |
+
title = {NAVSIM: Data-Driven Non-Reactive Autonomous Vehicle Simulation and Benchmarking},
|
| 43 |
+
author = {Daniel Dauner and Marcel Hallgarten and Tianyu Li and Xinshuo Weng and Zhiyu Huang and Zetong Yang and Hongyang Li and Igor Gilitschenski and Boris Ivanovic and Marco Pavone and Andreas Geiger and Kashyap Chitta},
|
| 44 |
+
booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
|
| 45 |
+
year = {2024},
|
| 46 |
+
}
|
| 47 |
+
```
|