Mistral enters embodied AI with Robostral Navigate, an 8B single-camera robot model

Mistral AI launched Robostral Navigate on July 8, its first robotics model — an 8-billion-parameter system that lets robots navigate complex environments using a single RGB camera and natural-language instructions.

The release marks the French lab’s move into embodied AI, extending it beyond text and image models into systems that control physical machines. Relying on one ordinary camera rather than lidar or multi-sensor rigs is the pitch: cheaper hardware for robots that follow spoken directions.

On benchmarks, Mistral says Robostral Navigate scores 76.6% success on the R2R-CE unseen-environment validation set and 79.4% on the seen-environment set, beating the best single-camera approach by 9.7 points and multi-sensor systems by 4.5 points. R2R-CE (Room-to-Room in Continuous Environments) is a standard test for language-guided navigation. The model was trained on roughly 400,000 simulated trajectories across 6,000 scenes and runs on wheeled, legged and flying robots, Mistral said; a later online reinforcement-learning stage added a further 3.2-point success-rate gain.

Those figures are the company’s own and were measured in simulation and benchmark settings, not independently verified, and navigation success in a benchmark does not guarantee reliability in a cluttered warehouse or a crowded hotel lobby. Mistral said the model targets manufacturing, delivery, logistics and hospitality, following recent deals with European industrial customers. Whether single-camera navigation holds up on real robots at those sites is the test that matters.