KE:SAI Open Science Autonomy Lab

Robot learning is bottlenecked by the cost of physical interaction. Our mission is to advance the efficiency frontier of robust & safe physical AI through fully open and reproducible research.

About KE:SAI

A Franco-German Partnership for Open Science

KE:SAI — Kyutai ELLIS Scalable Autonomous Intelligence — is a non-profit frontier AI research lab co-founded by kyutai and the ELLIS Institute Tübingen. Co-located in Tübingen and Paris, this Franco-German collaboration merges kyutai’s open science ethos and foundation model expertise with ELLIS’s talent attraction and foundational research excellence in 3D computer vision, data-driven simulation, causality and physical AI to pioneer world-leading research in world models and autonomy.

The Challenge

Physical AI is Inefficient and Closed

Physical AI is the next frontier of AI research. However, today's artificial intelligence is orders of magnitude less sample efficient than human intelligence, requiring very large datasets and compute resources for training. As robot learning is bottlenecked by the cost of physical interaction, Physical AI research is currently dominated by few proprietary silos with closed code, models and data. This hinders progress for the research field as a whole.

The Opportunity

Advancing the Research Frontier with Open Science

The open-source movement has been one of the most powerful forces in modern AI research. The release of models, datasets, and training recipes for large language models has triggered an explosion of innovation — compressing years of progress into months and enabling researchers worldwide to build on each other's work rather than starting from scratch. Physical AI has yet to benefit from this flywheel: robotic learning remains fragmented, with results that are difficult to reproduce, hardware-specific, and rarely shared in full. KE:SAI's mission is to democratize physical AI research by developing fully open stacks together with the research community. We are a small team of talented researchers, focused on this single mission. All our code and models will be released under permissive licenses that allow civil commercial use.

The Technology

Causal World Models

Our goal is to train policies that generalize effectively across a wide variety of embodiments and environments. To this end, we pioneer data- and compute-efficient methods to build foundation models for physical and causal AI. In contrast to current data-driven imitation learning approaches, we focus on hybrid, causal and latent world models, as well as Sim2Real techniques for integrating synthetic with real data. The hybrid nature of these world models allows for ingesting information from a variety of sources, including real-world data, simulations, and common sense knowledge from existing foundation models. Expressive and causally grounded latent spaces enable data- and compute-efficient closed-loop training of robot policies using (self-)supervised, reinforcement learning and self-play objectives.

Applications

A Fully Open Self-Driving Stack

Self-driving is the epitome of physical AI, covering multi-modal perception, planning in safety-critical situations, and control in highly dynamic multi-agent environments. Therefore, KE:SAI will first demonstrate the capabilities of the developed models by training self-driving policies using significantly less data and compute than typically required while reaching infraction rates competitive with frontier systems. Based on this foundation, KE:SAI will extend its effort to other areas of robotics.

The Team

Meet the Founders

We are a team of world-leading researchers with a proven track record of attracting talent. We

Received 27 best paper awards at leading vision, learning and robotics conferences

Introduced the KITTI benchmarks which transformed vision & robotics (>10M downloads)

Achieved top places at international self-driving competitions (nuPlan, CARLA, Waymo)

Open-sourced the two largest (1700h+) driving datasets (OpenDV, PhysicalAI-AV)

Developed the first reinforcement learning planner with state-of-the-art driving (CaRL)

Open-sourced the first large-scale video world model for driving (Vista)

Developed the first generative simulator for vectorized driving environments (SLEDGE)

Introduced pseudo-simulation for efficient closed-loop benchmarking (NAVSIM)

Pioneered methods for causal discovery from observational data

Triggered a paradigm shift in vision and graphics through work on implicit models

Established ELLIS and the most successful AI PhD program in Europe

Attracted top talent: former lab members co-founded Black Forest Labs, Mirelo AI and SpAItial AI

From left-to-right: Kashyap Chitta, Daniel Dauner, Andreas Geiger, Bernhard Schölkopf and Bernhard Jaeger

Thank You

Our Donors

KE:SAI would not exist without the vision and generosity of kyutai and its donors: the iliad Group, CMA CGM and Eric and Wendy Schmidt's philanthropy. Their...