Mistral has introduced Robostral Navigate, an 8-billion-parameter model developed specifically for embodied navigation. This model interprets RGB images and plain-language instructions to autonomously guide robots through complex real-world environments, including offices, residential buildings, commercial spaces, and even outdoor areas.
Unlike prior solutions that rely on depth sensors or multiple cameras, Robostral Navigate operates with a single RGB camera and achieves a 76.6% success rate on the R2R-CE benchmark for previously unseen environments. This performance surpasses the best single-camera approach by 9.7 points and outperforms leading multi-sensor systems by 4.5 points. The model runs across a diverse range of robots, including wheeled, legged, and aerial platforms, and is robust to variations in camera hardware.
The model was trained entirely in simulation, using approximately 400,000 navigation trajectories spanning 6,000 unique scenes. Mistral used a specialized data-generation pipeline and a tree-based attention-masking strategy for token-efficient supervised training, significantly reducing training tokens and shortening training time from months to days. The model also benefits from online reinforcement learning, which further increases its performance on navigation tasks.
Robostral Navigate is currently available to select partners in manufacturing, logistics, delivery, and hospitality sectors, with plans for broader access as testing continues. Robostral, the company behind this release, has focused on building advanced vision-language models and has designed this system entirely in-house, without relying on existing open-source models, highlighting its commitment to original research and application in robotic navigation.