Most self-driving AI systems today learn by analyzing how human drivers behave and then attempt to mimic those behaviors. This approach, known as imitation learning, works reasonably well in predictable traffic conditions but creates a critical blind spot: when the vehicle makes an unexpected or unsafe maneuver, the system often cannot explain why it chose that particular path over another. In an accident, that lack of transparency becomes a liability for regulators, insurers, and the public. A team at Seoul National University, led by professor Jun Won Choi, has developed a fundamentally different architecture called SafeDrive that solves this problem by evaluating multiple possible driving paths before the car even moves.
SafeDrive generates a set of candidate trajectories – each representing a potential future path for the vehicle – and then scores every one of them for safety using real-time sensor data. The model selects the trajectory with the highest safety score and executes it. This process, which the researchers call Fine-grained Safety Reasoning, creates a clear, auditable decision chain. If the car later crashes or violates traffic rules, engineers can review the scored alternatives to understand why the chosen path was deemed safest at that precise moment. This explainability is a major departure from the black-box nature of most current end-to-end autonomous driving systems.
The work was presented at the Conference on Computer Vision and Pattern Recognition (CVPR) 2026, the premier international conference in computer vision and AI. Out of thousands of submissions, only about 3% are selected as highlight papers – a designation reserved for the most significant contributions. SafeDrive earned that distinction, making it the first end-to-end autonomous driving paper developed by a Korean research team to receive a CVPR highlight. This achievement is a strong signal that South Korea is emerging as a serious player in AI-driven autonomy, a field long dominated by laboratories in the United States and China.
South Korea has made an enormous commitment to advanced technology: the government pledged 880 billion dollars over the next decade for investments in artificial intelligence, semiconductors, and robotics. SafeDrive is one of the first concrete academic results of that national push to gain recognition at the highest level. The model is not staying behind laboratory doors. It has already been integrated into EAD, a reference autonomous driving platform funded by the Korean Ministry of Trade, Industry and Energy. Professor Choi’s team is collaborating with domestic autonomous driving companies to test SafeDrive in real vehicles, using proprietary driving data collected on Korean roads. The plan is to push toward commercialization, with the goal of making autonomous vehicles not only safer but also more accountable.
The timing could not be more critical. Tesla’s robotaxi fleet in Austin, Texas, has been involved in accidents at a rate four times higher than that of human drivers. While Musk’s company defends its safety metrics, crashes involving autonomous vehicles raise urgent questions about accountability. When a human driver causes an accident, the legal and insurance system has well-established procedures to assign fault and understand the decision-making. For an autonomous system that relies on a neural network trained to copy human behavior, tracing the reasoning behind a split-second choice is extremely difficult. SafeDrive’s approach offers a solution: by explicitly scoring alternatives, it produces a decision trail that can be examined after an event. Regulators, courts, and insurers could use this information to determine whether the vehicle’s choice was reasonable given the circumstances.
The technical details of SafeDrive involve a multi-modal sensor fusion architecture that combines LIDAR point clouds, camera images, and radar data. The model simultaneously generates a set of plausible future trajectories using a variational inference method. Each trajectory is then evaluated by a safety assessment network that considers factors such as proximity to other objects, predicted intentions of other road users, lane boundaries, traffic signs, and dynamic constraints like braking limits. The scoring is fine-grained, meaning the model does not just output a binary safe/unsafe label but a continuous safety score that reflects a nuanced understanding of risk. The trajectory with the highest aggregate safety score is then executed by the vehicle’s low-level controller.
One of the key innovations is that the safety scoring is itself learned end-to-end from data, but the training objective is explicitly designed to rank trajectories by safety rather than just imitating human trajectories. This allows SafeDrive to potentially discover safer strategies that human drivers might not typically use, such as maintaining larger following distances in certain weather conditions or choosing a slightly longer path that avoids a blind intersection. The researchers demonstrated in simulated environments and on a real test vehicle that SafeDrive reduces the probability of collisions and near-misses compared to state-of-the-art imitation learning methods, while still maintaining efficiency comparable to human-like driving.
The broader context of autonomous driving development shows that the industry is increasingly realizing that end-to-end imitation learning has fundamental limitations. For example, Waymo and Cruise have traditionally used modular systems with separate perception, prediction, planning, and control components, each with explicit logic. That modularity aids safety analysis but can be brittle and slower. End-to-end systems, by contrast, are more fluid and can learn complex behaviors end-to-end, but they are notoriously hard to debug. SafeDrive sits in a middle ground: it is end-to-end in the sense that it learns from raw sensor data to produce a final action, but it introduces an intermediate scoring step that provides both transparency and a mechanism to directly optimize safety.
South Korea’s investment in this technology is already yielding downstream benefits. The integration with EAD means that the SafeDrive approach will be available to multiple Korean automotive companies and startups. Hyundai Motor Company, which has its own autonomous driving research division, could incorporate the methods into future production vehicles. The government’s strategy is to build a domestic ecosystem of hardware, software, and data to reduce reliance on foreign technology. With the CVPR highlight, SafeDrive also enhances the reputation of Korean AI research globally, which may attract international collaborations and talent.
Looking at the state of autonomous driving around the world, we see a race not only for performance but also for trust. Waymo operates commercial robotaxis in Phoenix and San Francisco, but those services have been plagued by incidents that caused vehicle stoppages, blocked traffic, and occasional collisions. In each case, Waymo publishes a safety report, but the opacity of end-to-end systems limits what can be shared. Regulators are pushing for more rigorous safety evaluation frameworks. The U.S. National Highway Traffic Safety Administration (NHTSA) has required companies to report crashes, but there is no standard for explaining decision-making. SafeDrive’s approach aligns with emerging regulatory trends: the European Union’s AI Act, for instance, requires high-risk AI systems to be transparent and traceable. An autonomous driving system that can show its reasoning would be better positioned to meet those requirements.
Beyond regulation, the insurance industry has a major stake in explainable autonomous driving. Insurers need to assess risk accurately to price policies for autonomous vehicles. A black-box system is essentially uninsurable if the insurer cannot understand why a crash occurred. SafeDrive’s decision trail could allow insurance companies to classify the severity of errors and adjust premiums accordingly. For fleet operators, such as ride-hailing companies, having a clear record of each decision can help in resolving disputes and improving safety over time.
The team at Seoul National University continues to refine the model. Current work focuses on extending the safety scoring to account for interactions between multiple autonomous vehicles, and handling rare edge cases such as pedestrians running into the road, animals, or construction zones. The researchers are also exploring how to embed ethical decision-making into the scoring function, so that when a collision is unavoidable, the system selects the trajectory that minimizes harm in a principled way. This is a notoriously difficult problem, and SafeDrive’s framework provides a natural way to incorporate such constraints because the scores can be weighted by a learned value function.
While SafeDrive is a landmark achievement for Korean autonomous driving research, it builds on a global body of work. Concepts similar to trajectory scoring have been explored in robotic planning for years, but SafeDrive is the first to demonstrate a full end-to-end system that learns the scoring function directly from sensor data and achieves state-of-the-art results on real-world benchmarks. The CVPR highlight committee recognized the novelty and potential impact. For the millions of people who will one day ride in autonomous vehicles, the hope is that such systems will make the roads safer and more trustworthy, one scored trajectory at a time.