DeepFusion AI

The problem

Cameras and LiDAR degrade in rain, fog, dust and darkness, and LiDAR remains costly for mass-market cars. In off-road tests reported by Byline Network (December 2025), LiDAR units failed within a minute in mud. Radar copes, but its raw output is too sparse and noisy for fine object detection.

The solution

RAPA (Real-time Attention-based Pillar Architecture) is perception software that fuses several 4D imaging radars in real time into a 360° view. It comes as radar-only (RAPA-R), radar plus LiDAR (RAPA-RL) or radar plus camera (RAPA-RC), on servers, edge computers or an ECU.

How it works

1. Several 4D radars (range, velocity, azimuth, elevation) around the vehicle capture point clouds, including in rain or darkness.
2. A pillar-based attention network built for sparse radar returns filters noise and separates static from moving objects.
3. Models train on a proprietary dataset plus about 2.5 years of simulation emulating several vendors' radars, which the company says yields a commercial model within three months of a new chip.
4. Outputs cover 3D detection, tracking and radar SLAM, with a claimed accuracy gain above 40% over competitors on a public benchmark.
5. Inference runs on embedded hardware, shipped as an ADAS fusion ECU or a module for surface drones and robots.

Value proposition

What the company calls LiDAR-level perception from cheaper radars, holding up in bad weather. For wave detection it claims over 90% accuracy in typhoons or night fog, up to 100 m ahead, more than 10 times per second. No cost savings figure published.

Technology edge

Most 4D radar players sell hardware; DeepFusion AI supplies vendor-agnostic software trained in simulation across several sensor families. RAPA handles radar point clouds natively instead of adapting camera or LiDAR networks, and claims over 40% higher accuracy on an unnamed public benchmark.

Business model

B2B: software licensing and module sales (fusion ECUs, USV, UGV, AMR), after pilots or PoCs. Pricing undisclosed.

Market

MarketsandMarkets (June 2025 report) projects the 4D imaging radar market to grow from 392.8 million dollars in 2025 to 1,206.9 million in 2030 (about 25% a year). Targets: ADAS and robotaxis, land and naval defense, shipping, logistics robotics.

Competitors and alternatives

Bitsensing (South Korea) builds its own 4D radars and raised a 25 million dollar Series B in June 2024. Mobileye said in May 2025 that a global automaker will use its imaging radar for eyes-off driving from 2028. Arbe (Israel), a key player per MarketsandMarkets, supplies 4D radar chipsets. DeepFusion AI sits as a software layer above such hardware.

Use cases

• A Chinese carmaker, Guangzhou Automobile (GAC), and ride-hailing operator DiDi were due to start pilot vehicles with the solution in February 2026, with 500 units planned by October 2026 (Byline Network, December 2025).
• A defense contractor, LIG Nex1, is working with DeepFusion AI on a reconnaissance unmanned surface vessel slated for delivery in early 2026, and the startup takes part in a combat USV program.
• A Japanese carmaker, Honda, planned a parking-space recognition proof of concept for early 2026, while PoC work with BMW and Volvo was under way or under discussion.

Traction and proof points

Seed funding from JNP Global (October 2024) while running PoCs in China, Korea and Japan. Exhibited at CES 2025 in the Incheon IFEZ pavilion. CES Innovation Awards 2026 Best of Innovation in Artificial Intelligence (CTA, November 5, 2025). Projects cited in late 2025: GAC and DiDi, LIG Nex1, PoCs with BMW, Volvo and Honda. Korean government Innovative Product label for wave detection. Series A in September 2026.

What remains to be proven

The 40% gain rests on an unnamed benchmark, and automotive customers are still at PoC or pilot stage with no public series contract. No public confirmation was found of the 500 GAC and DiDi units planned for October 2026. With funding amounts undisclosed, staying power through long automotive qualification cycles is hard to judge.