Yun Hong

Yun Hong

Computer Science student at Fudan University | Exchange student at UT Austin | Research in ML, Robotics, and Autonomous Driving

Learning Risk Map for Autonomous Driving in Partially Observable Environments

Published in IEEE Robotics and Automation Letters (RAL), 2025

Abstract

This research addresses the critical challenge of risk assessment in autonomous driving systems operating under partially observable conditions. We propose a comprehensive framework that combines advanced spatiotemporal modeling techniques with modern deep learning approaches to generate accurate risk maps for autonomous vehicle navigation.

Key Contributions

1. Risk Field Representation Engineering

  • Developed novel spatiotemporal modeling techniques for risk field representation
  • Engineered efficient data structures for real-time risk assessment
  • Integrated temporal dynamics with spatial risk distribution

2. Traffic Scene Generation

  • Implemented realistic traffic scene generation using diffusion models
  • Combined diffusion models with gradient optimization for enhanced realism
  • Generated diverse scenarios for comprehensive risk evaluation

3. Lightweight Neural Network Architecture

  • Designed and implemented a lightweight neural network for efficient risk prediction
  • Optimized for real-time performance in autonomous driving systems
  • Balanced accuracy and computational efficiency for practical deployment

Methodology

The proposed approach consists of three main components:

  1. Spatiotemporal Risk Modeling: Advanced techniques for capturing both spatial and temporal aspects of driving risks
  2. Diffusion-based Scene Generation: Novel application of diffusion models for generating realistic traffic scenarios
  3. Efficient Risk Prediction Network: Lightweight architecture optimized for real-time autonomous driving applications

Research Context

This work was conducted at MagicLab, Fudan University, under the guidance of Wenchao Ding from October 2024 to May 2025. The research focuses on gaining research experience and training in the field of autonomous driving and machine learning.

Status

Current Status: Preprint prepared and submitted to IEEE Robotics and Automation Letters (RAL) for peer review.

Impact

This research contributes to the field of autonomous driving by:

  • Providing a novel framework for risk assessment under partial observability
  • Demonstrating the effectiveness of diffusion models in traffic scene generation
  • Offering a practical solution for real-time risk prediction in autonomous vehicles

This research builds upon recent advances in:

  • Autonomous driving perception and planning
  • Diffusion models for scene generation
  • Risk assessment in robotics applications
  • Spatiotemporal modeling techniques

Technical Skills Applied

  • Deep Learning (PyTorch)
  • Computer Vision
  • Autonomous Driving Systems
  • Spatiotemporal Modeling
  • Diffusion Models
  • Neural Network Optimization

Recommended citation: Hong, Y., Ding, W. et al. (2025). "Learning Risk Map for Autonomous Driving in Partially Observable Environments." IEEE Robotics and Automation Letters (RAL). (Under Review)
Download Paper