AUTONEWS

A new AI model wants self-driving cars to think before they swerve

A research team led by Professor Jun Won Choi from the School of Electrical and Computer Engineering at Seoul National University has developed SafeDrive, an end-to-end autonomous driving AI model. The model employs a "fine-grained safety reasoning" approach, quantitatively evaluating the safety of multiple candidate driving trajectories based on perception outputs, ranking them accordingly, and selecting the optimal path to enhance both decision-making safety and interpretability. The SafeDrive model has been integrated into the EAD (Evolutionary Autonomous Driving) framework—a reference model for commercializing end-to-end autonomous driving spearheaded by Seoul National University with support from Korea’s Ministry of Trade, Industry and Energy. The research team is currently collaborating with Korean autonomous driving companies to conduct validation studies and advance real-world vehicle deployment of this technology. Choi stated that the team will continue optimizing the EAD model’s performance, aiming to achieve full-scale commercialization of end-to-end autonomous driving using larger and self-collected datasets. They also plan to build an open ecosystem to foster collaboration and knowledge sharing among academia, industry, and research institutions.

Most self-driving AI models study how humans drive and try to copy them. They work well in normal conditions but struggle to explain why they chose one path over another, which becomes a problem when a split-second decision goes wrong. A team at Seoul National University led by professor Jun Won Choi has built a model called SafeDrive that takes a different approach: it generates several possible trajectories, scores each one for safety using sensor data, and picks the path that scores best. The car shows its work.

The technique, called Fine-grained Safety Reasoning, was selected as a highlight paper at CVPR 2026, the leading computer vision and AI conference. Roughly 3% of submissions earn the distinction. It is the first time a Korean-developed end-to-end autonomous driving paper has received a CVPR highlight, a signal that South Korea is producing competitive research in a field dominated by US and Chinese labs. South Korea committed $880 billion over a decade to AI, chips, and robotics, and SafeDrive is one of the first results of that investment to earn top-tier academic recognition.

SafeDrive is not staying in the lab. It has been integrated into EAD, a reference model backed by Korea’s Ministry of Trade, Industry and Energy. Choi’s team is working with domestic autonomous driving companies to test it in real vehicles, with plans to push toward commercialisation using proprietary driving data. Tesla’s Austin robotaxis crash four times more than human drivers, illustrating that the safety and explainability problems SafeDrive addresses are not theoretical. When an autonomous vehicle makes a bad decision, regulators, insurers, and courts need to know why. A model that scores alternatives and selects the safest one produces an auditable decision trail that a black-box system cannot.

Self-driving cars have gotten pretty good at driving. The harder problem is teaching them to drive safely in situations nobody planned for. We have heard horror stories of self-driving cars, behaving erratically in emergency situations, often times delaying first responders from reaching the scene.

A team at Seoul National University, led by professor Jun Won Choi from the Department of Electrical and Computer Engineering, thinks they have cracked part of that puzzle with a new AI model called SafeDrive. The research was recently selected as a highlight paper at CVPR 2026, a distinction that goes only to roughly 3% of all submissions.

How does SafeDrive make driving decisions safer? Most end-to-end autonomous driving models work by studying massive amounts of real driving data and trying to mimic how humans react on the road. It works well most of the time, but these systems tend to struggle when it comes to explaining why they chose one path over another, and that becomes a real problem when safety is on the line.

Choi’s team built something called Fine-grained Safety Reasoning to fix this. Instead of picking one driving path and going with it, SafeDrive generates several possible trajectories, combines them with what the car’s sensors are perceiving, and scores each option for safety. The car then picks the path that scores best. It sounds simple, but it directly tackles the two biggest weaknesses of current end-to-end systems, safety and explainability.

Why is this such a big deal for Korea? As Tecplor reports, this is the first time a Korean-made end-to-end autonomous driving paper has landed a highlight spot at CVPR, one of the biggest AI and computer vision conferences in the world. It is a strong signal that Korea is no longer just watching from the sidelines while the US and China race ahead with their self-driving ambitions.

SafeDrive is not staying stuck in the lab either. It has already been folded into EAD, a reference model backed by Korea’s Ministry of Trade, Industry and Energy, and Choi’s team is now working with domestic autonomous driving companies to test it in real vehicles. Choi says the plan is to keep improving the model with bigger datasets and eventually push it toward full commercialization using their own collected data.

The new artificial intelligence model named SafeDrive, developed by researchers at Seoul National University led by Professor Jun Won Choi, calculates and scores the safety of multiple possible driving routes before the vehicle even begins to move. This cutting-edge technology addresses the two major challenges facing traditional autonomous vehicles: predictive safety in novel scenarios and the explainability of their decisions.

This breakthrough was hailed as a highlight of the prestigious CVPR 2026 computer vision conference.

How SafeDrive Works: Most AI systems for autonomous vehicles mimic human behavior based on historical data, leading to serious failures or unpredictable reactions in real-world emergencies. The SafeDrive model introduces an approach called Fine-grained Safety Reasoning:

Trajectory generation: Instead of blindly selecting a single path, the system instantly projects various route options.

Data integration: It combines these options in real-time with data from the car's sensors and cameras regarding its surroundings.

Quantitative scoring: The AI ​​assigns a safety score to each possible trajectory.

Optimized decision: The vehicle executes the route with the highest safety score.

Why this revolutionizes the market: Auditable trail: If the car needs to make a sudden maneuver, the AI ​​generates a clear record explaining why that specific route was chosen over others. This resolves a major legal issue for insurers and regulators.

Practical applications: SafeDrive has moved beyond academic labs and been integrated into the EAD (Evolutionary Autonomous Driving) framework, a benchmark model supported by South Korea's Ministry of Trade, Industry and Energy.

Real-world testing: The team is already validating the system on actual vehicles in partnership with Korean automakers and autonomous driving companies, aiming for full commercialization.

Source: School of Electrical and Computer Engineering at Seoul National University