The KIFLEX project, funded by the German Federal Ministry of Education and Research (BMBF) under grant 16ES1030, ran from 1 September 2019 to 31 August 2023. Its aim was to create a highly flexible, low‑power, multi‑core deep‑learning ASIC that can be integrated into autonomous vehicles and still execute future, unknown neural‑network concepts. The hardware platform is programmable and reconfigurable, allowing it to adapt to different operating states of the data‑processing system depending on the surrounding environment and the data stream. This adaptability is a core innovation, as it enables the same chip to run a wide range of perception algorithms without redesign.

Central to the project is the fusion of LiDAR point clouds with camera imagery, a technique that meets the current needs of VDA automation levels 4 and 5. Neural‑network‑based algorithms perform sensor‑signal evaluation and early fusion, providing a highly accurate environmental model that underpins subsequent decision making. The software framework developed in parallel supports real‑time calibration and map‑based localisation, ensuring that the fused data can be used immediately for vehicle control. Together, the hardware and software form a tightly coupled system that delivers high performance while keeping power consumption low, a balance that is essential for automotive deployment.

The Fraunhofer Institute for Open Communication Systems (FOKUS) contributed a full scientist position, focusing on the design, implementation, and validation of perception neural networks to run on the ASIC, the development of the innovative fusion algorithm, and the conception, implementation, and testing of the sensor‑data‑fusion framework. Three universities, a small‑ and medium‑sized enterprise (KMU), and a large industry partner joined the consortium, each bringing complementary expertise. The collaboration combined academic research, industrial know‑how, and the Fraunhofer system‑engineering capability to move the technology from concept to a prototype ready for commercial integration.

By integrating algorithm design with hardware development, KIFLEX demonstrates a holistic approach that is rare in the autonomous‑driving field. The resulting platform can execute complex deep‑learning workloads in real time, enabling new algorithms that were previously too demanding for on‑board processors. The project’s outcomes are expected to feed into the commercial products of the partner companies, accelerating the availability of high‑performance, low‑power AI accelerators for autonomous vehicles. The KIFLEX effort thus represents a significant step toward making advanced sensor‑fusion and perception algorithms commercially viable for the next generation of self‑driving cars.