Valeo has partnered with KE, a non-profit research lab founded by Kyutai and the ELLIS Institute Tübingen, to develop a generalist artificial intelligence driving system and release its code, trained models and datasets openly.
The project aims to establish a reproducible reference system that automakers, automotive suppliers and researchers can use to develop, evaluate and benchmark automated-driving technologies. The partners plan to match and eventually surpass leading results on major public driving benchmarks.
The initiative addresses the limited availability of open driving models. According to the partners, most automated-driving systems used in production remain proprietary, restricting automakers’ ability to inspect, modify and independently reproduce the underlying software.
Developing a Generalist Driving Model
The collaboration brings together Valeo’s Paris-based research laboratory, valeo.ai, and the Tübingen-based KE team, which has experience with driving datasets and benchmarks including KITTI and NAVSIM.
Valeo’s research work includes the DrivoR and VaVAM models. The new project aims to build on this research to develop a compact model capable of operating across different vehicles, sensor configurations and driving environments.
The partners intend to make the resulting system and its development resources available to the wider research and automotive communities. This approach would allow external teams to examine the technology, reproduce experiments and compare their own systems against a common reference.
Training With Real-World Data and Simulation
The development process will begin with training on real-world driving data. The partners then plan to use simulation-based reinforcement learning and learned world models to expose the system to a wider range of driving situations.
Simulation is intended to help address a limitation of recorded driving data: rare and safety-critical events may occur too infrequently in real-world datasets to provide sufficient examples for training.
By generating and exploring additional scenarios, the research team aims to improve the model’s ability to respond to varied conditions. However, performance on research benchmarks alone would not establish that a system is ready for deployment in production vehicles.
Marc Vrecko, CEO of Valeo Brain, said:
This open-science collaboration with KE marks a strategic milestone for Valeo and the broader European automotive ecosystem. By combining our deep industrial and technological expertise in autonomous driving with world-class open-source research, we are establishing a shared foundation to accelerate the development of physical AI and democratize next-generation self-driving technologies.
Andreas Geiger, CEO of KE, said:
For language models, open source compressed years of progress into months, yet frontier physical AI is still largely closed. Valeo shares our belief that open foundations create more innovation than any closed system ever could. Everything we develop will be open. Every step forward is a step forward for everyone.
Open Research and Automotive Independence
The project could provide European automakers and suppliers with greater access to the underlying technology used in automated-driving development. Open code, model weights and datasets can allow organizations to evaluate systems independently rather than relying exclusively on externally controlled software platforms.
For European manufacturers, this approach may also support efforts to maintain greater control over strategic automotive technologies. Access to open models could give research teams more flexibility to adapt software to their own vehicles, sensors and computing platforms.
Valeo’s participation also connects open AI research with an established automotive supplier’s industrial expertise. While the model layer would be made available openly, companies would still need to develop and validate the hardware, integration and vehicle-level systems required for commercial use.
Benchmark Performance Is Only One Step
The partners’ objective to match and surpass leading public benchmarks provides a measurable target for the research. Benchmarks such as NAVSIM can help researchers compare driving-system performance under standardized conditions.
However, benchmark results do not replace production validation. Automakers adopting open models would still need to assess system reliability, performance across operating conditions, functional safety, regulatory compliance and liability before deploying the technology in vehicles.
The collaboration’s practical impact will therefore depend not only on its benchmark results but also on the quality and usability of its public releases, the extent of independent adoption and the ability of developers to translate research results into validated automotive systems.
The partners have positioned the initiative as an open foundation for further research and development. Specific release dates, model specifications and benchmark targets have not been detailed in the supplied announcement.
