Mistral Bets on Physical AI With 8 Bn Parameter Autonomous Robot Navigation Model

French AI startup Mistral has expanded beyond large language models into robotics with the launch of Robostral Navigate, its first embodied AI model designed for robots to navigate complex environments using only a single RGB camera and natural language instructions. 

The 8-billion-parameter model allows wheeled, legged, and aerial robots to interpret verbal instructions such as "go to the kitchen" or "move to the loading dock" while autonomously planning their route through unfamiliar environments. 

Unlike conventional robotic navigation systems that rely on LiDAR, depth sensors, or multiple cameras, Robostral Navigate operates using a single standard RGB camera, significantly reducing hardware complexity and deployment costs. 

According to Mistral, the model achieves a 76.6% success rate on the unseen R2R-CE benchmark and 79.4% on seen environments, outperforming the best existing monocular navigation approach by 9.7 percentage points and surpassing systems that use depth sensors or multiple cameras by 4.5 percentage points. The company said the model was trained entirely in simulation before being transferred to real-world robotic platforms. 

The company said Robostral Navigate is hardware-agnostic, enabling it to work across different robot manufacturers and form factors without requiring specialised sensor configurations. Mistral positions the model as a foundation for logistics, warehouse automation, industrial inspection, and other enterprise robotics applications where cost-effective navigation is critical. 

The launch follows Mistral's acquisition of Austrian robotics startup Emmi AI in May, signalling the company's broader ambitions in physical AI.