Arthur Mensch's Mistral Enters Physical AI: Robostral Navigate Steers Robots From a Single RGB Camera

Mistral AI·medium signal

On July 8 Mistral released Robostral Navigate, an 8B, hardware-agnostic navigation model that lets robots traverse offices, homes, and outdoor sites using one RGB camera plus language prompts — hitting 76.6% on unseen R2R-CE, beating the best multi-sensor systems while being cheaper. It was trained entirely in simulation (~400K trajectories across 6,000 scenes) then refined with online RL. The signal for builders: a frontier LLM lab is porting its stack into embodied AI, and sim-only-plus-RL training is now producing single-camera policies that out-perform depth/multi-cam rigs.