Add robotics pipeline tag and sample usage
Browse filesThis PR enhances the model card for **CoIRL-AD** by:
- Adding the `pipeline_tag: robotics` to the metadata, which will improve discoverability on the Hugging Face Hub for models related to autonomous driving.
- Incorporating the "Quick Start" section from the project's GitHub repository as a "Sample Usage" section, providing users with immediate instructions on how to run the model for inference.
The existing paper, code, and project page links are maintained via their respective badges at the top of the model card.
Please review and merge if these changes are appropriate.
README.md
CHANGED
|
@@ -1,36 +1,65 @@
|
|
| 1 |
-
---
|
| 2 |
-
license: apache-2.0
|
| 3 |
-
|
| 4 |
-
|
| 5 |
-
|
| 6 |
-
|
| 7 |
-
|
| 8 |
-
<
|
| 9 |
-
<a href="https://github.
|
| 10 |
-
<a href="https://github.
|
| 11 |
-
<a href="https://github.com/
|
| 12 |
-
<a href="https://
|
| 13 |
-
<a href="https://
|
| 14 |
-
<a href="https://
|
| 15 |
-
<
|
| 16 |
-
|
| 17 |
-
|
| 18 |
-
|
| 19 |
-
|
| 20 |
-
|
| 21 |
-
|
| 22 |
-
<
|
| 23 |
-
<a href="https://github.
|
| 24 |
-
<a href="https://
|
| 25 |
-
|
| 26 |
-
|
| 27 |
-
</div>
|
| 28 |
-
|
| 29 |
-
|
| 30 |
-
|
| 31 |
-
|
| 32 |
-
|
| 33 |
-
|
| 34 |
-
|
| 35 |
-
|
| 36 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
pipeline_tag: robotics
|
| 4 |
+
---
|
| 5 |
+
|
| 6 |
+
# CoIRL-AD: Collaborative–Competitive Imitation–Reinforcement Learning in Latent World Models for Autonomous Driving
|
| 7 |
+
|
| 8 |
+
<div align="center">
|
| 9 |
+
<a href="https://seu-zxj.github.io/">Xiaoji Zheng</a>*,
|
| 10 |
+
<a href="https://ziyuan-yang.github.io">Yangzi Yuan</a>*,
|
| 11 |
+
<a href="https://github.com/Ian-cyh">Yanhao Chen</a>,
|
| 12 |
+
<a href="https://github.com/doraemonaaaa">Yuhang Peng</a>,
|
| 13 |
+
<a href="https://github.com/TANGXTONG1">Yuanrong Tang</a>,
|
| 14 |
+
<a href="https://openreview.net/profile?id=~Gengyuan_Liu1">Gengyuan Liu</a>,
|
| 15 |
+
<a href="https://scholar.google.com/citations?user=_Wrx_yEAAAAJ">Bokui Chen</a>‡ and
|
| 16 |
+
<a href="https://scholar.google.com/citations?user=AktmI14AAAAJ">Jiangtao Gong</a>‡.
|
| 17 |
+
<div>
|
| 18 |
+
*: Equal contribution.
|
| 19 |
+
‡: Corresponding authors.
|
| 20 |
+
</div>
|
| 21 |
+
|
| 22 |
+
<div>
|
| 23 |
+
<a href="https://seu-zxj.github.io/CoIRL-AD"><img alt="Static Badge" src="https://img.shields.io/badge/_-page-blue?style=flat&logo=githubpages&logoColor=white&logoSize=auto&labelColor=gray"></a>
|
| 24 |
+
<a href="https://arxiv.org/abs/2510.12560"><img alt="Static Badge" src="https://img.shields.io/badge/arxiv-paper-red?logo=arxiv"></a>
|
| 25 |
+
<a href="https://github.com/SEU-zxj/CoIRL-AD"><img alt="Static Badge" src="https://img.shields.io/badge/github-code-white?logo=github"></a>
|
| 26 |
+
<a href="https://huggingface.co/Student-Xiaoji/CoIRL-AD-models"><img alt="Static Badge" src="https://img.com/img.shields.io/badge/hf-models-yellow?logo=huggingface"></a>
|
| 27 |
+
</div>
|
| 28 |
+
|
| 29 |
+
</div>
|
| 30 |
+
|
| 31 |
+
**CoIRL-AD** introduces a dual-policy framework that unifies imitation learning (IL) and reinforcement learning (RL) through a collaborative–competitive mechanism within a latent world model.
|
| 32 |
+
The framework enhances generalization and robustness in end-to-end autonomous driving without relying on external simulators.
|
| 33 |
+
|
| 34 |
+
|
| 35 |
+
|
| 36 |
+
Here we provide our model checkpoints (see `/ckpts`), info files (see `/info-files`) for dataloader to download and reproduce our experiment results.
|
| 37 |
+
|
| 38 |
+
## Sample Usage (Quick Start)
|
| 39 |
+
To run our code quickly, the best way is to try run the inference code, see the results and visualize the results.
|
| 40 |
+
1. download ckpt of CoIRL-AD via this [link](http://huggingface.co/Student-Xiaoji/CoIRL-AD-models/blob/main/ckpts/CoIRL-AD/epoch_21.pth), or ckpt of LAW we trained on our machine via this [link](https://huggingface.co/Student-Xiaoji/CoIRL-AD-models/blob/main/ckpts/LAW/epoch_15.pth).
|
| 41 |
+
2. run inference code.
|
| 42 |
+
|
| 43 |
+
2.1 open the config file `./projects/configs/coirl.py`
|
| 44 |
+
|
| 45 |
+
2.2 set results save path at line 110-111:
|
| 46 |
+
```
|
| 47 |
+
save_results_flag=True,
|
| 48 |
+
results_path="results/coirl.pth"
|
| 49 |
+
```
|
| 50 |
+
|
| 51 |
+
2.3 inference
|
| 52 |
+
```bash
|
| 53 |
+
# evaluate CoIRL-AD on nuscenes eval set
|
| 54 |
+
./tools/dist_test.sh <path of configuration file coirl.py> <path of CoIRL-AD ckpt> <num-of-GPUs>
|
| 55 |
+
# evaluate LAW on nuscenes eval set (you can skip)
|
| 56 |
+
./tools/dist_test.sh <path of configuration file law.py> <path of LAW ckpt> <num-of-GPUs>
|
| 57 |
+
```
|
| 58 |
+
3. visualize the results.
|
| 59 |
+
3.1 the most simple way is to directly visualize the trajectory in `coirl.pth`, if you want to generate visualized figs and videos in our paper, see following steps.
|
| 60 |
+
3.2 open `./projects/tools/visualization/vis_result_compare.py`
|
| 61 |
+
3.3 configure path of `data_root`, `info_path`, and `output_result_path` (here you can ignore `baseline_result_path`)
|
| 62 |
+
3.4 run the visualization script
|
| 63 |
+
```bash
|
| 64 |
+
python path-of-vis_result_compare.py --save-path <directory-to-save-visualization-results>
|
| 65 |
+
```
|