[ICRA 2026] Conformal Risk Tube Prediction

Project Page | Paper

_{Risk Tube Prediction models the joint spatial–temporal uncertainty of risk.}

This repository contains the official codebase for training, evaluation, and visualization of the methods described in:

Uncertainty-Aware Vision-based Risk Object Identification via Conformal Risk Tube Prediction
Kai-Yu Fu and Yi-Ting Chen
National Yang Ming Chiao Tung University

_{Visual Risk Object Identification Result.}

_{Conformal Calibration Result.}

⚙️ Getting Started

📝 System Setup

• Operating System: Linux Ubuntu 18.04
• Python Version: 3.7
• PyTorch Version: 1.10.1
• CUDA Version: 11.3
• GPU: Nvidia RTX 3090
• CPU: Intel Core i7-11700KF

📥 Dependency Installation

1. Clone the Repository

   git clone https://github.com/doraemonhappy/my-first-repo.git

2. Create and activate a new Conda environment:

   cd CRTP
   conda env create -f environment.yml --name CRTP
   conda activate CRTP

📦 Datasets Downloads

_{Multiple Coexisting Risks Dataset}

We construct the Multiple Coexisting Risks dataset, which integrates the four risk categories including Interaction, Collision, Obstacle, and Occlusion. Within a single scenario, multiple risk categories can occur concurrently or in sequence. In total, we obtain about 1,000 scenarios, which enables comprehensive validation under multi-risk settings.

• Download Multiple_Coexisting_Risks_Dataset here. Please extract all train{xx}.zip files and place their contents into the same folder.
• Please refer to the dataset description for more details.

🚀 Usage

Training

# step 1: Risk Category Classifier Pre-training
python train_cls.py

# step 2: Full Architecture CRTP Training
python train.py

Evaluation (Metric: Coverage, TV, TC, BA, Risk-IOU)

# Risk Category Classifier Inference
python inference_cls.py

# Full Architecture CRTP Inference
python inference.py               # Given GT Bounding Box 
python inference_yolo_detector.py # Given Perception Bounding Box 

Visualization & Downstream Task (Braking Alerts)

# Visualization and Save Braking Alerts (data root folder should contain a single scenario)
python vis_roi_and_save_braking_alerts.py --mode 'vis_save'

# Braking Alerts Metric Evaluation (data root folder should contain all scenarios)
python vis_roi_and_save_braking_alerts.py --mode 'metric'

Visualization of Braking Alerts

# First run CRTP and every baseline in ./Baselines/Braking_Alerts to save all braking alerts.
python compare_all_roi_braking_alerts.py

_{Our calibrated, temporally aligned Risk Tube suppresses nuisance interventions.}

📄 BibTeX

If our work contributes to your research, please consider citing it with the following BibTeX entry:

TBD

🙌 Acknowledgment

We acknowledge that the implementation used in this project are adapted from RiskBench, SAOCP.

Thanks to these great open-source projects!