Website | Huggingface | arXiv | Model | Discord
- [2024/08/27] We update the latest results under the new protocols with two new metrics and fixed bugs. The older ones are deprecated with Bench2Drive version before commit 311c35d294c1d4e5350e4b8adcc682c28b9556a8 and Bench2DriveZoo version before commit 31432e868c3ca1bef5c7fa39ba4bd4e7a3e7538a.
- [2024/08/19] [Major Updates] To better assess driving performance, we add two additional metrics: Driving Efficiency and Driving Smoothness. Consequently, we remove the penalty for minimum speed in calculating the Drive Score and extend the TickRunTime from 2000 to 4000 to allow for a more lenient driving evaluation. We are currently reassessing all baselines. Please stay tuned.
- [2024/08/10] We update the team_code agent of UniAD and VAD to fix the camera projection bug mentioned in 2024/07/29. Their corresponding scores will be uploaded soon with new metrics introduced.
- [2024/07/29] As kindly suggested in an issue, there is a bug in the team code agent of UniAD and VAD during evaluation, i.e, the BACK CAMERA's extrinsic is wrong. The training process is correct. To be consistent with the reported results, we do not modify the code right now. Users' are strongly encouraged to use the correct extrinsics.
- [2024/07/22] We add more reminders in the evaluation code to avoid the miss of logs. According to Haochen's kind suggestion, we add automatic cleaning code in the evaluation toolkit. Users' may set in their bash script to restart the evaluation infinitely until finishing the evaluation since CARLA is easy to crash.
- [2024/07/10] We further clean and add more clips in the Full set (13638 clips). Since HuggingFace only allows up to 10000 files per repo, we use two repos to store the Full set. As suggested in this issue issue, we add a filelist and sha256 of clips for each set.
- [2024/06/19] Due to a typo in the upload script for HuggingFace, all clips of scenario VehicleTurningRoutePedestrian are empty in the HuggingFace version. We have fixed that. Please check your data to make sure they are not empty.
- [2024/06/05] Bench2Drive realases the Full dataset (10000 clips), evaluation tools, baseline code, and benchmark results.
- [2024/04/27] Bench2Drive releases the Mini (10 clips) and Base (1000 clips) split of the official training data.
- The datasets has 3 subsets, collected by our strong world model based RL expert Think2Drive, namely Mini (10 clips), Base (1000 clips) and Full (10000 clips), to accommodate different levels of computational resource.
- Detailed explanation of dataset structure, annotation information, and visualization of data.
| Subset | Hugging Face |
Baidu Cloud |
Approx. Size | File List |
|---|---|---|---|---|
| Mini | Download script | - | 4G | Mini Json File |
| Base | Hugging Face Link | Baidu Cloud Link | 400G | Base Json File |
| Full | Full HF Link - 9888 files/Sup HF Link - 3814 file | - | 4T | Full/Sup Json File |
Note that the Mini Set is 10 representative scenes. You may download them by manually select file names from the Base set.
Use the command line: huggingface-cli download --repo-type dataset --resume-download rethinklab/Bench2Drive --local-dir Bench2Drive-Base to download from hugginface. User may consider mirror site if Huggingface is blocked. Use BaiduPCS-Go to download from Baidu Cloud. Both command lines are resumable.
- Download and setup CARLA 0.9.15
mkdir carla cd carla wget https://carla-releases.s3.us-east-005.backblazeb2.com/Linux/CARLA_0.9.15.tar.gz tar -xvf CARLA_0.9.15.tar.gz cd Import && wget https://carla-releases.s3.us-east-005.backblazeb2.com/Linux/AdditionalMaps_0.9.15.tar.gz cd .. && bash ImportAssets.sh export CARLA_ROOT=YOUR_CARLA_PATH echo "$CARLA_ROOT/PythonAPI/carla/dist/carla-0.9.15-py3.7-linux-x86_64.egg" >> YOUR_CONDA_PATH/envs/YOUR_CONDA_ENV_NAME/lib/python3.7/site-packages/carla.pth # python 3.8 also works well, please set YOUR_CONDA_PATH and YOUR_CONDA_ENV_NAME
- Add your agent to leaderboard/team_code/your_agent.py & Link your model folder under the Bench2Drive directory.
Bench2Drive\ assets\ docs\ leaderboard\ team_code\ --> Please add your agent HEAR scenario_runner\ tools\ --> Please link your model folder HEAR
- Debug Mode
# Verify the correctness of the team agent， need to set GPU_RANK, TEAM_AGENT, TEAM_CONFIG bash leaderboard/scripts/run_evaluation_debug.sh - Multi-Process Multi-GPU Parallel Eval. If your team_agent saves any image for debugging, it might take lots of disk space.
# Please set TASK_NUM, GPU_RANK_LIST, TASK_LIST, TEAM_AGENT, TEAM_CONFIG, recommend GPU: Task(1:2). # It is normal that certain model can not finsih certain routes, no matter how many times we restart the evaluation. It should be treated as failing as it usually happens in the routes where agents behave badly. bash leaderboard/scripts/run_evaluation_multi_uniad.sh
- Visualization - make a video for debugging with canbus info printed on the sequential images.
python tools/generate_video.py -f your_rgb_folder/
- Metric: Make sure there are exactly 220 routes in your json. Failed/Crashed status is also acceptable. Otherwise, the metric is inaccurate.
# Merge eval json and get driving score and success rate # This script will assume the total number of routes with results is 220. If there is not enough, the missed ones will be treated as 0 score. python tools/merge_route_json.py -f your_json_folder/ # Get multi-ability results python tools/ability_benchmark.py -r merge.json # Get driving efficiency and driving smoothness results python tools/efficiency_smoothness_benchmark.py -f merge.json -m your_metric_folder/
- CARLA has complex dependencies and is not stable. Please check the issue section of CARLA very carefully.
- Use tools/clean_carla.sh frequently and multiple times. Some CARLA processes are difficult to kill. Be sure to clean_carla could avoid lots of bugs.
- In our evaluation tools, the launch of CARLA is automatic: https://github.com/Thinklab-SJTU/Bench2Drive/tree/main/leaderboard/leaderboard/leaderboard_evaluator.py#L203. But you could always start CARLA by the one single command line to debug.
- CARLA is not controlled CUDA_VISIBLE_DEVICES! It is controlled by -graphicsadapter in the command line. Interestingly, in some machines, for some unknown reasons, -graphicsadapter=1 is not available. For example, with 4 GPUS, it might be: GPU0 -graphicsadapter=0, GPU1 -graphicsadapter=2, GPU2 -graphicsadapter=3, GPU3 -graphicsadapter=4.
- The conflict of PORT is frequently happened. Use lsof-i:YOUR_PORT frequently to avoid conflict. Avoid use small port numbers (<10000 could be unsafe).
- 4.26.2-0+++UE4+Release-4.26 522 0 Disabling core dumps. Only showing these two lines without termination is good. WARNING: lavapipe is not a conformant vulkan implementation, testing use only. is bad.
- Re-install vulkan might be helpful sudo apt install vulkan-tools; sudo apt install vulkan-utils
- We find that nvidia driver version 470 is good all the time. 515 has some problems but okay. 550 has lots of bugs.
- sleep is important to avoid crash of CARLA. For example, https://github.com/Thinklab-SJTU/Bench2Drive/blob/main/leaderboard/leaderboard/leaderboard_evaluator.py#L207, the sleep time should be extended for slower machines. When it comes to multi-gpu evaluation, https://github.com/Thinklab-SJTU/Bench2Drive/blob/main/leaderboard/scripts/run_evaluation_multi_uniad.sh#L58, the sleep time should also be extended for slower machines.
----------------------The results below are deprecated due to change of protocols and bugs----------------------
All assets and code are under the Apache 2.0 license unless specified otherwise.
Please consider citing our papers if the project helps your research with the following BibTex:
@article{jia2024bench,
title={Bench2Drive: Towards Multi-Ability Benchmarking of Closed-Loop End-To-End Autonomous Driving},
author={Xiaosong Jia and Zhenjie Yang and Qifeng Li and Zhiyuan Zhang and Junchi Yan},
journal={arXiv preprint arXiv:2406.03877},
year={2024}
}
@inproceedings{li2024think,
title={Think2Drive: Efficient Reinforcement Learning by Thinking in Latent World Model for Quasi-Realistic Autonomous Driving (in CARLA-v2)},
author={Qifeng Li and Xiaosong Jia and Shaobo Wang and Junchi Yan},
booktitle={ECCV},
year={2024}
}