metadata
task_categories:
- robotics
Fail2Drive: Benchmarking Closed-Loop Driving Generalization
Project Page | Paper | GitHub
Fail2Drive is the first CARLA v2 benchmark designed to test closed-loop generalization on truly unseen long-tail scenarios. By pairing each shifted route with an in-distribution reference scenario, it exposes substantial hidden failure modes in current state-of-the-art driving models.
Highlights
- 17 unseen scenarios for evaluation of true generalization.
- 30 novel assets including animals, visual noise, and adversarial obstacles.
- Paired route design enables quantification of the generalization gap.
- 100 route pairs in diverse environments and configurations.
- Toolbox for creating custom obstacles and routes.
Installation
To set up the environment and the Fail2Drive CARLA simulator, follow these steps:
# 1. Clone this repository
git clone https://github.com/autonomousvision/fail2drive.git
cd fail2drive
# 2. Set up the Fail2Drive CARLA simulator
mkdir f2d_carla
curl -L \
https://huggingface.co/datasets/SimonGer/fail2drive/resolve/main/fail2drive_simulator.tar.gz \
| tar -xz -C f2d_carla
# 3. Create the conda environment
conda env create -f environment.yml
conda activate fail2drive
# 4. Set environment variables
source env_vars.sh
Sample Usage
To run a keyboard-controlled human agent on a benchmark route, start CARLA in a separate terminal and run:
python leaderboard/leaderboard/leaderboard_evaluator.py \
--agent ${WORK_DIR}/leaderboard/leaderboard/autoagents/human_agent_keyboard.py \
--routes ${WORK_DIR}/fail2drive_split/Generalization_PedestriansOnRoad_1085.xml
To run the PDM-Lite expert policy:
python leaderboard/leaderboard/leaderboard_evaluator_local.py \
--agent ${WORK_DIR}/team_code/visu_agent.py \
--track MAP \
--routes ${WORK_DIR}/fail2drive_split/Generalization_PedestriansOnRoad_1085.xml
Citation
@article{gerstner2024fail2drive,
title={Fail2Drive: Benchmarking Closed-Loop Driving Generalization},
author={Gerstner, Simon and others},
journal={arXiv preprint arXiv:2604.08535},
year={2024}
}