French AI company Mistral has released its first AI model for robot navigation, Robostral Navigate, with a total of 8B parameters. This model allows robots to achieve full autonomous navigation in complex environments using only a single regular RGB camera, without relying on depth sensors or LiDAR.

image.png

The model is mainly designed for embodied navigation tasks, and its application scenarios cover offices, homes, commercial buildings, and outdoor environments. Traditional robot navigation usually requires LiDAR or depth sensors to perceive the surrounding space, which involves high hardware costs and deployment complexity. Robostral Navigate significantly lowers this barrier, enabling a complete closed-loop from environmental perception to path planning with just a regular camera and an 8B model.

Outperforming "multi-eye" with a single eye, success rate in new scenarios exceeds 76%

The performance data is impressive. In the R2R-CE benchmark test, the model achieved a success rate of 79.4% in scenes already present in the training set, and 76.6% in completely new scenes. More notably, it outperforms the previous best single-camera solution by 9.7 points, and even surpasses the best systems using depth sensors or multiple cameras by 4.5 points. This means that the "single-eye" solution not only matches the "multi-eye" solution but also achieves a comprehensive breakthrough.

The model was fully developed internally by Mistral, and was trained exclusively in simulated environments, using approximately 400,000 recorded paths spread across 6,000 different virtual spaces. This pure simulation training strategy greatly reduces reliance on real-world data collection, while also verifying the effectiveness of transferring training from virtual environments to real-world scenarios.

Compatible with wheeled, legged, and flying robots, open-source ecosystem is promising

In terms of compatibility, Robostral Navigate is suitable for three types of robots: wheeled, legged, and flying, covering the current mainstream robot forms. From wheeled transport robots in warehousing and logistics, to quadrupedal robotic dogs, to drones, the same navigation model can be adapted, demonstrating strong versatility.