Valeo and KE:SAI Launch Open-Source Franco-German Project for General-Purpose Self-Driving AI

Valeo, the Paris-listed automotive supplier, and KE:SAI, a non-profit open-science lab co-founded by Paris-based Kyutai and the ELLIS Institute Tübingen, announced on October 13 a joint research programme to build a self-driving system based on general-purpose, fully reproducible AI. Everything the partners produce, including code, trained models and datasets, will be released as open source, according to their announcement, carried by Zonebourse.

The stated aim is to match and then beat the performance of the main public driving benchmarks. The teams plan a single AI model able to drive different vehicles with different sensors in varied environments. Training will start on large volumes of real driving recordings, with an architecture built to adapt to different camera and sensor set-ups. Reinforcement learning in simulation and learned world models will then expose the model to rare, safety-critical situations that are scarce or missing in real-world data.

The pairing brings together valeo.ai, the AI research lab of Valeo, known for open models and papers such as DrivoR and VaVAM, and the founding team of KE:SAI, behind open benchmarks and driving architectures including KITTI, NAVSIM and TransFuser. KE:SAI was set up in May 2026 with a founding donation from Kyutai and works from Tübingen and Paris.

What it means in practice

The partners say nearly all driving models deployed in the car industry are proprietary, leaving manufacturers and suppliers dependent on technology they cannot inspect, adapt or reproduce. A shared open reference would let carmakers, suppliers and researchers reproduce results, build on the model and compare their own systems against a common baseline. Marc Vrecko, chief executive of Valeo Brain, called the collaboration a strategic step for Valeo and the European automotive ecosystem. Andreas Geiger, chief executive of KE:SAI, said open source let language models make years of progress in months, while cutting-edge physical AI remains largely closed.

The bet on openness echoes the open-weight release of Mistral Large 4 and the evaluation work in the 13.5 million euro AI safety programme launched by Inria. What to watch: when the first models and datasets are published, how they score on public benchmarks, and whether other European carmakers adopt the shared baseline.