DRIVE: Dependable Robust Interpretable Visionary Ensemble Framework in Autonomous Driving

This paper was accepted at ICRA 2025 (IEEE International Conference on Robotics and Automation 2025, CCF B, Core A*). This repository is the official PyTorch Lightning implementation.

Paper (PDF)	Project Page	Code
arXiv	xll0328.github.io/drive	GitHub

Authors: Songning Lai~, Tianlang Xue~, Hongru Xiao, Lijie Hu, Jiemin Wu, Ninghui Feng, Runwei Guan, Haicheng Liao, Zhenning Li, Yutao Yue†

---

Overview

DRIVE is a comprehensive framework for improving the dependability and stability of explanations in end-to-end unsupervised autonomous driving models. It addresses instability issues in concept-based explainable systems (e.g. built on Concept Gridlock (DCG)) and provides both consistent/stable interpretability and consistent/stable control output.

• Consistent interpretability — explanations align with model behavior across samples.
• Stable interpretability — explanations remain robust under input perturbations.
• Consistent output — predictions are reliable across scenarios.
• Stable output — performance is robust under perturbations.

Teaser: see the project page for the animated overview. Framework figure:

---

Table of Contents

• Method
• Installation
• Project structure
• Setup
• Arguments
• Training & evaluation
• Analysis & scripts
• Citation
• License

---

Method

DRIVE defines four key attributes and implements a training/evaluation pipeline to achieve them:

1. Consistent interpretability — concept explanations are aligned with the model’s actual reasoning.
2. Stable interpretability — concept explanations remain stable under input noise/perturbations.
3. Consistent output — driving decisions are consistent across similar inputs.
4. Stable output — driving performance is robust under perturbations.

The codebase builds on a concept-bottleneck driving model: first train a DCG-style checkpoint, then resume with the DRIVE pipeline (see Training & evaluation) to optimize for the above properties. Novel metrics for evaluating dependability are included in the analysis notebooks.

---

Installation

git clone https://github.com/xll0328/ICRA25-DRIVE.git
cd ICRA25-DRIVE
conda create -n drive python=3.8
conda activate drive
pip install -r DRIVE/requirement.txt

---

Project structure

ICRA25-DRIVE/
├── README.md
├── LICENSE
├── DRIVE/
│   ├── requirement.txt
│   ├── dataloader_comma.py
│   ├── comma_preprocess/          # Data preprocessing (raw_readers, clean_up_data)
│   ├── DCG_Core/                  # Core DCG model and training
│   │   ├── main.py, model.py, module.py
│   │   └── README.md
│   └── DRIVE/
│       ├── main_copy1.py          # Main entry for DRIVE training/testing
│       ├── model_copy1.py, model_copy2.py, module_copy1.py
│       ├── losses.py, PGD.py
│       ├── sh_scripts/            # Training and evaluation scripts
│       ├── attacks/               # Perturbation and noise tools
│       ├── scenarios_DCG/         # Scenario configs
│       └── analysis/              # Notebooks for MAE, concept visualization, performance

---

Setup

1. Dataset

• Comma2k19: commaai/comma2k19
• Download: Academic Torrents

2. Data preprocessing

Extract video frames, obtain sample data, and save to HDF5:

• Run DRIVE/comma_preprocess/raw_readers.ipynb
• Run DRIVE/comma_preprocess/clean_up_data.ipynb

Paths inside the notebooks and scripts (e.g. dataset path, checkpoint path) should be set to your local environment.

---

Arguments

Argument	Example	Description
-task	angle / distance / multitask	Task for the model
-train	True	Enable training
-test	False	Enable testing
-gpu_num	1	Number of GPUs
-dataset_path	/path/to/dataset	Path to dataset and checkpoints
-concept_features	True	Use concept bottleneck
-new_version	True	Train from scratch (do not resume)
-save_path	/path/to/savepath	Path for saving CSV predictions
-num-gpus	1	Number of GPUs
-max_epochs	200	Maximum epochs
-bs	4	Batch size
-checkpoint_path	path/to/checkpoint.ckpt	Checkpoint for resume / test
img_noise	GaussianNoise	Noise type (build noisy dataset via DRIVE/DRIVE/attacks/ if needed)

---

Training & evaluation

1. Train a DCG checkpoint (without DRIVE)

Follow Concept Gridlock (DCG) to obtain a DCG checkpoint.

2. Resume with DRIVE

From the repository root (or from DRIVE/DRIVE/ if your paths are set there):

python main_copy1.py -dataset comma -backbone none -concept_features -ground_truth normal \
  -train -gpu_num 1 -max_epochs 50 -task distance -bs 2 -checkpoint_path DCG_checkpoint

More examples: see DRIVE/DRIVE/sh_scripts/ (e.g. run_experiments_concept_comma_DRIVE.sh, testconcept.sh).

---

Analysis & scripts

• MAE and predictions: DRIVE/DRIVE/mae_test.ipynb — load checkpoints, run prediction, compute MAE and save results.
• Concept visualization: DRIVE/DRIVE/analysis/showconcepts.ipynb — visualize concept activations and compare across conditions.
• Model performance: DRIVE/DRIVE/analysis/observe_model_performances.ipynb — aggregate metrics and plots.
• Perturbation / noise: DRIVE/DRIVE/attacks/ — e.g. noise-img.py for adding Gaussian noise to HDF5 image data; adjust h5_file_path and output paths for your setup.

---

Citation

@inproceedings{lai2025drive,
  title={DRIVE: Dependable Robust Interpretable Visionary Ensemble Framework in Autonomous Driving},
  author={Songning Lai and Tianlang Xue and Hongru Xiao and Lijie Hu and Jiemin Wu and Ninghui Feng and Runwei Guan and Haicheng Liao and Zhenning Li and Yutao Yue},
  booktitle={Proceedings of the IEEE International Conference on Robotics and Automation (ICRA)},
  year={2025},
  eprint={2409.10330},
  archivePrefix={arXiv},
  primaryClass={cs.RO},
  url={https://arxiv.org/abs/2409.10330},
}

---

License

This project is licensed under the MIT License — see LICENSE.

References

• Concept Gridlock (DCG)
• Comma2k19