This Korean AI Lets Cars See the Road Using Only Radar

Every self-driving car on the road today uses the same expensive combination: cameras, LiDAR, and radar working together, each covering the other's weak spots. LiDAR alone can cost tens of thousands of dollars per vehicle. The assumption has always been that you need all three to drive safely.

Deep-fusion AI built a system that changes that equation.

The South Korean startup, founded in Incheon in 2022 by CEO SungHun Yu, developed RAPA — Real-time Attention-based Pillar Architecture — a perception system that reads 4D radar signals alone to build a complete picture of the road around a moving vehicle. No cameras required. No LiDAR. Just radar, and deep learning that understands what the radar is seeing.

At CES 2026, RAPA won the Best of Innovation Award in Artificial Intelligence — the top recognition in AI at the world's largest consumer tech show. By October 2026, a global mobility company is expected to begin mass production of autonomous taxis using RAPA technology.

Why Autonomous Driving Stayed Expensive for So Long

The cost of self-driving systems has slowed adoption more than the technology itself. LiDAR — the laser-based sensor that gives vehicles precise 3D spatial awareness — was for years the most accurate perception tool available. It was also prohibitively expensive for mass-market vehicles.

Camera systems brought costs down but introduced new problems. Rain, fog, glare, and darkness degrade camera performance exactly when reliable perception matters most. The industry response was to combine all three sensors, letting each compensate for the others. The approach worked but kept costs high and software complexity even higher.

Radar was historically limited — good at detecting objects at a distance but unable to tell you much about what they were or where they sat in three dimensions. Then 4D imaging radar changed that. A newer generation of radar hardware that adds height data to the traditional range, velocity, and angle measurements, producing dense point clouds that approach what LiDAR generates, at a fraction of the cost.

How RAPA Works in Plain Terms

Most perception systems treat radar as a backup — a secondary check after cameras and LiDAR have already processed the scene. RAPA does something different. It takes raw radar signals and feeds them directly into a deep learning model, making the radar itself the primary perception source rather than a supporting one.

Think of it this way: instead of translating what the radar sees into a format that matches what cameras and LiDAR see, RAPA teaches the AI to understand radar data on its own terms — as its own language. The result is a system that doesn't need the other sensors to make sense of the world.

Using four corner radars positioned around a vehicle, RAPA builds 360-degree environmental awareness in real time — mapping the road, identifying objects, and tracking movement continuously as the vehicle drives. With five 4D imaging radars running simultaneously at 288 channels, DFAI achieved 300,000 valid data points per second — enough resolution to accurately read complex urban environments.

The underlying architecture, which DFAI calls the Perceptive Sensor Standard, defines perception as a consistent structure rather than a fixed set of sensors. That matters practically: if an automotive OEM switches to a different radar hardware supplier, only a short fine-tuning step is needed. The perception logic stays the same. This makes RAPA adaptable across different vehicle configurations without rebuilding the model from scratch.

DFAI also solved a major data bottleneck. Training AI perception models typically requires enormous real-world datasets — millions of annotated driving miles. DFAI built a virtual radar simulation environment instead, modeling real RF sensors digitally to generate synthetic radar signals for pre-training. The model learns in simulation, then transfers to real hardware with minimal additional training. This dramatically cuts development time and cost compared to camera or LiDAR-based systems that need physical data collection at scale.

Where This Is Already Going

DFAI isn't just pitching a concept. A global mobility company begins mass production of autonomous taxis using RAPA by October 2026 — three years after the company was founded. That timeline, from founding to mass production contract, is rare in a sector known for long development cycles and slow commercialization.

Domestically, DFAI has integrated its technology into an unmanned surface vehicle project in collaboration with a Korean defense partner — the same radar-first approach applied to autonomous ships navigating coastal and maritime environments where cameras struggle with fog and low visibility.

The company is also developing RAPA-RC, which adds early-fusion with camera data, and RAPA-RL, which adds LiDAR fusion — both built on the same Perceptive Sensor Standard. These aren't replacements for the radar-first approach; they're extensions of it, adding sensor options without changing the underlying perception structure. New functions and expanded use cases are planned for CES 2027.

DFAI received support through Incheon Startup Park and the Incheon Free Economic Zone — part of a Korean startup ecosystem that sent 14 companies to CES 2026, earning 17 Innovation Awards between them.

CEO SungHun Yu described the company's starting point plainly: not at the frontier of what technology can do, but at the point where existing systems start to fail — fog, rain, glare, sensor cost — and where the real question isn't "what is possible" but "what can actually be sustained across real-world conditions."