github.com/Thinklab-SJTU/Bench2Drive @main sqlite

NeurIPS 2024 Datasets and Benchmarks Track

Website | Huggingface | arXiv | Model | Discord

What can Think2Drive + Bench2Drive provide ? Please click to view the video. ↓↓↓

Table of Contents:

1. News
2. Dataset
3. Benchmark
4. License
5. Citation

News

• [2026/05/21] Check Bench2Drive-Robust, which analyze models' performance under real deployment issues: camera-stream failures, ego-state estimation errors, and compute-control latency.
• [2026/03/27] Check Bench2Drive-Speed, which estimates a new & useful functionality of AD systems - speed customization!
• [2025/10/13] Check Bench2Drive-VL, which enables closed-loop QA of VLM4AD. It also provides a efficient communication interface between VLM environment (python 3.9 or higher) and CARLA (only available for python 3.7 and 3.8). Users encount Python Version error should consider the protocol as well.
• [2025/02/18] In our latest work DriveTransformer (ICLR25), a tiny validation set Dev10 is proposed for quick development of models. The 10 clips are carefully selected from the official 220 routes, to be both difficult and representative with low variance. It is suggested to be used for ablation study to avoid overfitting the whole bench2drive220 routes.
• [2025/02/05] As meticulously described in CARLA_GARGE, L2 error is not a meaningful indictor at all. We agree that authors' should stop reporting results on nuScenes open-loop planning and reviewers should not ask for nuScenes open-loop planning results. The number on nuScenes open-loop planning is not persuasive at all and could only lead to unjustified and wrong conclusions, which may impede the development of the field. As a member of the community, we call for stopping using nuScenes open-loop planning as well.
• [2024/10/14] As kindly pointed out in an issue, typos were discovered in the ability calculation. We have corrected the typos and updated the multi-ability results. This update will not affect driving score and success rate.
• [2024/10/14] As kindly pointed out in an issue, a bug was discovered in the B2D_vad_dataset. We leave it here to be consistent with existing results. We do not anticipate any major influence caused by this bug. Users may fix the bug according to their needs.
• [2024/09/26] Bench2Drive is accepted at NeurIPS 2024 Datasets and Benchmarks Track.
• [2024/08/27] We update the latest results under the new protocols with two new metrics and fixed bugs.
• [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.
• [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.

Dataset

• 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.

Student Model Code (with Think2Drive as Teacher Model)

• Uniad/VAD in Bench2Drive
• TCP/ADMLP in Bench2Drive

Setup

• Download and setup CARLA 0.9.15 bash 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

Eval Tools

• Add your agent to leaderboard/team_code/your_agent.py & Link your model folder under the Bench2Drive directory. bash Bench2Drive\ assets\ docs\ leaderboard\ team_code\ --> Please add your agent HEAR scenario_runner\ tools\ --> Please link your model folder HEAR
• Debug Mode bash # 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. bash # 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. bash 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. ```bash # 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/
  

  ```

  Deal with CARLA

• 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. Showing these two lines without termination is good. WARNING: lavapipe is not a conformant vulkan implementation, testing use only. could be good while sometimes it has bugs.
• If you face issues, always try to start CARLA in one single line to make sure CARLA could run. If CARLA is finished immediately, it is very possible to be related to Vulkan. Try /usr/bin/vulkaninfo | head -n 5
• Re-install vulkan might be helpful sudo apt install vulkan-tools; sudo apt install vulkan-utils In the end, you need to make sure your vulkan is correct. We have tested Vulkan Instance Version: 1.x WARNING: lavapipe is not a conformant vulkan implementation, testing use only. and version 1.1/1.2/1.3 works fine.
• 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.

Benchmark

• V0.0.3(Currently in use)

• Fix typos in the ability calculation. (Issue #112)

• V0.0.2(Depracated)

• Fix the BACK CAMERA's extrinsic bug. (Issue #36)
• Turn up tickruntime (2000 -> 4000).
• Drive Score removes min speed penalty.

• Code Version

  • Bench2Drive
  • Bench2DriveZoo

• V0.0.1(Depracated)

• Initial Version
• Code Version
  • Bench2Drive
  • Bench2DriveZoo

License

All