TFv5 / README.md
nielsr's picture
nielsr HF Staff
Improve model card for Fail2Drive and TransFuser++
aaac843 verified
|
raw
history blame
2.37 kB
metadata
language:
  - en
library_name: pytorch
license: cc-by-4.0
model_name: TransFuser++ (TFv5) CARLA Garage Checkpoints
pipeline_tag: robotics
tags:
  - carla
  - carla-simulator
  - autonomous-driving
  - imitation-learning
  - end-to-end-driving
  - carla-leaderboard
  - transfuser
  - transfuser++

TransFuser++ (TFv5) CARLA Garage Checkpoints

This repository contains pretrained checkpoints for the TransFuser++ (TFv5) model, which serves as a baseline for the Fail2Drive benchmark.

Fail2Drive: Benchmarking Closed-Loop Driving Generalization

Project Page | Code

Introduction

Fail2Drive is a paired-route benchmark for closed-loop generalization in CARLA, featuring 200 routes and 17 new scenario classes. It highlights failure modes in state-of-the-art models when facing distribution shifts in appearance, layout, behavior, and robustness. TransFuser++ is an end-to-end autonomous driving agent that utilizes both LiDAR and camera inputs.

Sample Usage

To run the TransFuser++ model within the Fail2Drive environment, follow the installation instructions in the official repository.

Once the environment and simulator are set up, you can evaluate the agent using the following command:

LIVE_VISU=1 python leaderboard/leaderboard/leaderboard_evaluator_local.py \
  --routes ${WORK_DIR}/fail2drive_split/Generalization_PedestriansOnRoad_1085.xml \
  --agent ${WORK_DIR}/team_code/sensor_agent.py \
  --agent-config /path/to/checkpoint_folder

Note: The --agent-config parameter should point to the directory containing both model.pth and config.json.

Attribution

This repository re-uploads the pretrained checkpoints originally released by the CARLA Garage project. All credit and attribution belong to the original authors.

Original source: CARLA Garage Models

Citation

@article{ger2024fail2drive,
  title={Fail2Drive: Benchmarking Closed-Loop Driving Generalization},
  author={Ger, Simon and others},
  journal={arXiv preprint arXiv:2604.08535},
  year={2024}
}