Are We Ready for Snow-tonomous Driving? Challenges Toward Robust Automated Driving in Extreme Weather

Eren Erdal Aksoy, Abu Mohammed Raisuddin, Thien-Thu Ngo, Eren Aydemir, Osman Tahir Ekşi, Paula Litkey, Petri Manninen, Heikki Hyyti, Jyri Maanpää, Tarmo Kivioja, Tuomas Herranen, Väisänen Sofie, Niemi Kalle, Isabelle Morvan, Pierre Duthon, Ian Marsh, Martin Sanfridson, Henrik Toss, Thanh Bui, Yuri Poledna, Maikol Funk Drechsler, Mikko Tarkiainen, Matti Kutila, Topi Miekkala, Stefan Parzefall, Valentina Donzella, Pak Hung Chan, Sogol Kharrazi, Mario CeccarelliarXiv 2026 · arXiv

TL;DR

The ROADVIEW project synthesizes four years of large-scale research to identify ten critical challenges preventing autonomous vehicles from mastering extreme weather. The paper argues that achieving "snow-tonomous" driving requires a holistic systems-level shift—moving from static, clear-weather models to adaptive, multimodal architectures that can quantify and act upon environmental uncertainty in real-time.

What Problem Does This Paper Address?

While autonomous driving (AD) has made massive strides in sunny climates like California or Arizona, harsh weather remains its "final frontier." Heavy snow, rain, and fog don't just reduce visibility for human drivers; they physically occlude sensors with ice, introduce significant noise into LiDAR point clouds, and fundamentally change vehicle dynamics through reduced road friction.

Most current AD stacks are "context-agnostic," meaning they use fixed strategies for sensor fusion and control that fail when a camera is blinded by a blizzard or a road becomes a sheet of ice. Furthermore, there is a massive "data desert" for these conditions; most public datasets contain less than 2% snowy scenes, making it nearly impossible to train robust AI models for winter environments.

Methods

The authors present a comprehensive framework developed under the EU-funded ROADVIEW project. Key technical approaches include:

  • Adaptive Multimodal Fusion: Instead of static weighting, they propose "weather-aware" fusion (e.g., the LRC-WeatherNet architecture). This system dynamically adjusts the importance of different sensors—relying more on RADAR and Thermal cameras when RGB cameras and LiDAR are degraded by snow.
  • Early Denoising: Implementing supervised and unsupervised point cloud filtering to remove "snowflake noise" before data reaches downstream perception tasks.
  • EA-NDT Positioning: An "Environment-Aware Normal Distributions Transform" for HD map-based localization. It uses semantically segmented LiDAR data to maintain centimeter-level accuracy even when lane markings are completely buried.
  • Uncertainty-Aware Control: A velocity controller that translates spatially variable estimates of road grip and visibility into actionable safety margins and braking distances.
Fig. 1: ROADVIEW demonstration scene in heavy snowy conditions in Finland, where the minimum recorded temperature during testing was -34°C (Click here to see the demonstration video). The top-left image shows the onboard camera and LiDAR sensors from a test vehicle, both heavily covered with snow and ice.

Datasets

The paper introduces several new benchmarks to address the scarcity of harsh-weather data:

  • Snowy Scenes: The first multimodal dataset collected in real-world snowy conditions using 128-beam LiDAR, RGB, and Thermal cameras with dense point-wise annotations for snowflakes.
  • M3 Scenes: A massive 800km dataset spanning multiple seasons and environments, including LiDAR, RADAR, and RGB sensors.
  • REHEARSE: A unique "paired frame" dataset (identical scenes captured with and without controlled noise) designed specifically to validate the fidelity of weather noise models in simulation.
Fig. 2: Sample images from the ROADVIEW Snowy Scenes [63] (left) and M3 Scenes [64] (right) datasets. Both datasets have different multimodal sensor setups and provide dense point-wise annotations, including falling snowflakes, and 3D bounding boxes to benchmark 3D semantic segmentation, object detection, and denoising under harsh weather conditions.

Key Insights and Ideas

  • Denoising is a "Free" Performance Boost: Early filtering of weather-induced noise in raw sensor data can improve semantic segmentation by over 5% mIoU and object detection by 2% mAP without requiring any changes to the underlying perception models.
  • The "Sim-to-Real" Gap is Granular: In noise modeling, the drop-size distribution of rain or snow often affects sensor data more than the overall weather intensity. Standard simulators often oversimplify this, leading to "brittle" models.
  • V2X Needs Quality Metadata: Collective perception (sharing data between cars and infrastructure) is vital in snow, but it's useless unless messages include "confidence indicators." A car needs to know why a roadside sensor is reporting a detection—is it a clear view or a noisy, low-confidence guess?
  • Anticipatory vs. Reactive Safety: By using look-ahead friction estimation, vehicles can move from reactive braking (slipping then correcting) to anticipatory safety (braking earlier on ice vs. snow).
Visualization of weight maps in the adaptive gated fusion mechanism, showing how the system shifts reliance between sensors in clear vs. snowy conditions.

Key Results

  • Positioning Robustness: The EA-NDT system provided better orientation accuracy than traditional GNSS/INS in challenging winter environments, proving that LiDAR-based HD maps are a viable fallback when satellite signals are unreliable.
  • Real-time Weather Classification: The LRC-WeatherNet demonstrated the ability to classify weather types and road conditions in real-time, allowing the vehicle to adjust its "safe operating envelope" on the fly.
  • Validation of Noise Models: Using the REHEARSE dataset, the team successfully validated physics-based noise models for LiDAR and 4D RADAR, narrowing the gap between virtual testing (XiL) and real-world performance.

Limitations and Future Work

Despite these advances, several hurdles remain:

  • Regulatory Fragmentation: Inconsistent EU regulations regarding liability (who is at fault if a V2X-guided car crashes?) and data privacy (GDPR vs. camera-based sensing) create significant commercial barriers.
  • Standardization Gaps: There is currently no industry-wide consensus on metrics to quantify "data quality" or "simulation credibility" for adverse weather.
  • Transparency (XAI): Most perception models remain "black boxes." For safety-critical winter driving, there is an urgent need for Explainable AI that can justify its decisions to human auditors.
  • Computational Cost: High-fidelity, weather-aware models are computationally expensive, making real-time execution on edge hardware a continuing engineering challenge.

Summary updated

Abstract

Automated driving has gained remarkable momentum over the past decade, with both industry and academia investing substantial resources in developing perception and control solutions to advance vehicle autonomy. However, these efforts have largely focused on clear-weather scenarios, leaving harsh conditions, such as snow, rain, and fog, comparatively understudied. As a result, inclement weather remains one of the most significant hurdles for autonomous vehicles, often leading to substantial performance degradation in downstream perception and control tasks. In this paper, we present insights from the EU-funded Horizon Europe project ROADVIEW, which aims to develop robust, weather-aware perception and control modules for connected and automated vehicles operating under severe conditions, including snow, rain, and fog. Drawing on four years of large-scale development and real-world demonstrations across multiple vehicle platforms in different locations, we identify ten key challenges that hinder the development of weather-resilient autonomy, including data scarcity, domain shifts, limitations in multimodal perception, control robustness, and validation gaps, among others. We summarize these challenges and lessons learned toward enabling reliable snow-autonomous driving, highlighting critical directions for future research and standardization.
BibTeX
@misc{aksoy2026are,
  author = {Eren Erdal Aksoy and Abu Mohammed Raisuddin and Thien-Thu Ngo and Eren Aydemir and Osman Tahir Ekşi and Paula Litkey and Petri Manninen and Heikki Hyyti and Jyri Maanpää and Tarmo Kivioja and Tuomas Herranen and Väisänen Sofie and Niemi Kalle and Isabelle Morvan and Pierre Duthon and Ian Marsh and Martin Sanfridson and Henrik Toss and Thanh Bui and Yuri Poledna and Maikol Funk Drechsler and Mikko Tarkiainen and Matti Kutila and Topi Miekkala and Stefan Parzefall and Valentina Donzella and Pak Hung Chan and Sogol Kharrazi and Mario Ceccarelli},
  title = {{Are We Ready for Snow-tonomous Driving? Challenges Toward Robust Automated Driving in Extreme Weather}},
  year = {2026},
  eprint = {2611.15217},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.15217},
}
Figure