Valeo and KE:SAI will open-source a generalist autonomous driving model
The French supplier and a Franco-German non-profit research lab said they would jointly train one driving model and release its code, trained weights and datasets as open source.
Valeo and KE:SAI, a non-profit research lab co-founded in May 2026 by kyutai and the ELLIS Institute Tübingen, said on October 13 that they would jointly train a single generalist autonomous driving model and release every component of it — code, trained models and datasets — as open source.
The two described a compact model designed to transfer across different sensor and camera configurations rather than being rebuilt for each one. It would be trained on large-scale real-world driving recordings, then extended with reinforcement learning in simulation and with learned world models covering rare and safety-critical situations. Both sides brought existing open-source work to the effort: Valeo listed DrivoR and VaVAM from its valeo.ai research unit, and KE:SAI listed the KITTI dataset, the NAVSIM benchmark and the TransFuser model.
Valeo set the initiative against the state of the industry, saying nearly every driving model deployed in the automotive sector is proprietary and leaves automakers “dependent on technology they cannot inspect, adapt or reproduce.” Marc Vrecko, chief executive of Valeo Brain, and Andreas Geiger, chief executive of KE:SAI, were named in the release.
Neither party disclosed funding, headcount or a term. Valeo is listed in Paris, and the French central register of regulated information held no filing from the company after an October 8 voting-rights declaration covering September, and none at all from KE:SAI or kyutai — so the absence holds against the compelled record and not only against the two press releases.
Source: Valeo newsroom