Skip to main content

Mistral's new 8B model lets robots navigate with just a regular camera

French AI company Mistral has released its first model for robot navigation, Robostral Navigate. With 8 billion parameters, it lets robots navigate complex environments using just a single regular RGB camera—no depth sensors or LiDAR required.

Image

The model is built for embodied navigation tasks in offices, homes, commercial buildings, and outdoor spaces. Traditional robot navigation often relies on expensive LiDAR or depth sensors to perceive surroundings. Robostral Navigate cuts that cost and complexity, creating a complete loop from perception to path planning with just a camera and the model.

Outperforming multi-camera systems

The performance numbers are striking. In the R2R-CE benchmark, the model achieved a 79.4% success rate in familiar scenes and 76.6% in entirely new ones. That's 9.7 points higher than the previous best single-camera solution, and even 4.5 points above the best systems using depth sensors or multiple cameras. In other words, a single camera can now beat setups with multiple eyes.

Mistral developed the model entirely in-house, training it exclusively in simulated environments using about 400,000 recorded paths across 6,000 virtual spaces. This pure simulation approach reduces reliance on real-world data collection and shows that skills learned in virtual worlds transfer effectively to real ones.

Works with wheeled, legged, and flying robots

Robostral Navigate is compatible with three robot types: wheeled, legged, and flying. That covers everything from warehouse transport robots and quadrupedal robotic dogs to drones. The same navigation model adapts to all, showing strong versatility.

The open-source nature of the model also promises a thriving ecosystem. Developers can integrate it into their own robots without needing expensive hardware, potentially accelerating progress in robotics.

Key Points

  • Mistral's Robostral Navigate is an 8B-parameter model for autonomous robot navigation using a single RGB camera.
  • It outperforms multi-camera and depth-sensor systems in the R2R-CE benchmark, with over 76% success in new scenes.
  • The model was trained entirely in simulation on 400,000 paths across 6,000 virtual environments.
  • It works with wheeled, legged, and flying robots, offering broad compatibility.