French artificial intelligence company Mistral AI has expanded beyond large language models into robotics with the launch of Robostral Navigate, an AI model designed to enable autonomous robots to navigate complex environments using just a single RGB camera and natural language instructions.
The announcement marks Mistral's first step into embodied AI, a rapidly growing field focused on enabling AI systems to interact with and operate in the physical world. Unlike many robotic navigation systems that rely on LiDAR, depth sensors or multiple cameras, Robostral Navigate uses a standard RGB camera while delivering competitive navigation performance.
According to Mistral AI, the 8 billion parameter model can interpret spoken or written instructions such as directing a robot to move through corridors, enter rooms and stop at designated locations. The company said the model achieved a 76.6% success rate on the unseen validation benchmark of the Room-to-Room in Continuous Environments (R2R-CE), outperforming the best existing single camera approach by 9.7 percentage points and surpassing leading multi sensor systems by 4.5 percentage points.
The company said Robostral Navigate has been built entirely in house and does not rely on existing open source vision language models. Instead, it is based on Mistral's proprietary visual grounding technology, which enables the model to identify objects, understand scenes and determine where a robot should move next. Navigation capabilities were then developed using approximately 400,000 simulated trajectories generated across 6,000 virtual environments.
A key element of the system is its pointing based navigation approach. Rather than issuing traditional movement commands based solely on distance or coordinates, the model predicts the exact location in the robot's camera view where it should move next. If the destination falls outside the current field of view, the system switches to conventional movement instructions until the target becomes visible again. Mistral said this approach improves adaptability across different robot types and camera configurations.
The company also introduced a token efficient training technique based on prefix caching, allowing complete navigation episodes to be processed in a single forward pass without information leakage between time steps. Combined with reinforcement learning, this reduced training complexity while improving the model's ability to adapt to real world environments that differ from simulated training scenarios.
Robostral Navigate has been designed to operate across wheeled, legged and aerial robots, making it suitable for applications in warehouses, manufacturing facilities, logistics centres, commercial buildings and hospitality environments. Mistral said the model can navigate dynamic spaces containing people and unexpected obstacles without requiring expensive sensor hardware.
The launch comes as technology companies increasingly invest in physical AI, where robotics is becoming a strategic extension of generative AI. Advances in foundation models are enabling robots to combine perception, reasoning and navigation capabilities, creating opportunities for greater automation across industrial and commercial sectors.
Industry observers view Mistral's latest release as a significant expansion of the company's AI portfolio, which has until now focused primarily on language, coding and enterprise productivity models. By entering robotics, Mistral joins a growing group of AI developers exploring embodied intelligence as the next frontier for artificial intelligence.
The company said Robostral Navigate represents the first step in its long term robotics strategy, with further research planned to improve autonomous navigation and enable robots to operate more effectively across diverse real world environments. Mistral is also expanding its robotics research team as it continues development in embodied AI.