Insider Brief
- Mistral AI introduced Robostral Navigate, an 8B robot-navigation model that uses RGB images and plain-language instructions to move robots through environments.
- Mistral said the model uses a single RGB camera without LiDAR, depth sensors or multiple cameras, and achieved a 76.6% success rate on R2R-CE validation unseen.
- The company said Robostral Navigate was trained entirely in simulation using about 400,000 trajectories across 6,000 scenes and improved further through online reinforcement learning.
Mistral AI has introduced Robostral Navigate, an 8B model built for robot navigation. The model takes RGB images and plain-language instructions and uses them to move a robot through their environment, according to Mistral.
“Our model is designed for robotic navigation, enabling robots to autonomously navigate complex environments, including offices, residential and commercial buildings, and outdoor settings,” the company noted in a blog post announcing the new model.
According to Mistral,, Robostral Navigate uses a single RGB camera and does not require LiDAR, depth sensors or multiple cameras, according to Mistral. The company said the model achieved a 76.6% success rate on R2R-CE validation unseen, a benchmark for following navigation instructions in environments not used during training.
That result was 9.7 points higher than the best single-camera approach and 4.5 points higher than the best system using depth or multiple cameras. The company reported the model also achieved a 79.4% success rate on validation seen, according to the company.
The model is intended to support robotic navigation in manufacturing, delivery, logistics and hospitality. Mistral said it can run on wheeled, legged and flying robots and can generalize across robot sizes.
How Robostral Works
Robostral Navigate predicts where a robot should move next by pointing to image coordinates in the robot’s current camera view and estimating the desired orientation at arrival. Mistral said this approach makes the model more robust to differences in camera intrinsics and world scale.
When the target location is outside the camera’s current field of view, the model can instead use local movement commands, such as moving forward or sideways and turning by a specified amount.
Mistral indicated Robostral Navigate was built in-house and does not rely on existing open-source vision-language models. The model was initialized from its own vision-language model for grounding tasks, including pointing, counting and object localization.
Training the Model
The company said it trained the model entirely in simulation using a data-generation pipeline that produced about 400,000 trajectories across 6,000 scenes. It also used a training method based on prefix-caching, which reduced the number of training tokens by 22 times compared with training on one sample per time step. The company said that turned training runs that would have taken months into runs completed in days.
After supervised training, Mistral said it used online reinforcement learning through CISPO to improve the model’s performance. The company said that stage helped the model learn from trial and error, recover from failures and improve exploratory behavior, raising the success rate by 3.2%.
Mistral said Robostral Navigate is its first step and that it is expanding its robotics team.
“We believe navigation is a foundational capability for general-purpose robotics,” the company noted. “By combining large-scale simulation, efficient training, and strong grounding priors, Robostral Navigate demonstrates that state-of-the-art embodied navigation can be achieved with a compact model and a single RGB camera.”
