Mistral AI Launches Robostral Navigate for Single-Camera Robot Navigation
Mistral AI has released Robostral Navigate, its first AI model for robot navigation. The 8-billion-parameter model allows robots to navigate complex environments autonomously using only a single RGB camera.
PARIS – Mistral AI announced on July 8 the launch of Robostral Navigate, an artificial intelligence model designed for robot navigation. The model enables robots to autonomously navigate complex environments, such as offices, homes, and outdoor areas, relying solely on a single RGB camera and without the need for depth sensors or LiDAR.
The 8-billion-parameter model is Mistral AI's initial offering for embodied navigation tasks, allowing robots to perceive and maneuver through their surroundings. The company stated that the model was developed internally and trained exclusively in simulated environments, utilizing approximately 400,000 paths across 6,000 distinct virtual spaces.
In the R2R-CE benchmarking test, Robostral Navigate achieved a success rate of 79.4% in 'seen' scenarios and 76.6% in 'unseen' scenarios, which represent environments not encountered during training. This performance marks an improvement of 9.7 percentage points over previous best single-camera solutions and 4.5 percentage points over systems employing depth sensors or multiple cameras.
Mistral AI reports that Robostral Navigate is suitable for various robot types, including wheeled, legged, and flying platforms. Training solely within simulation aims to streamline development and reduce the need for extensive physical testing, facilitating broader application across different robotic systems.