Mistral has introduced Robostral Navigate, an 8B model designed for autonomous robotic navigation. The model enables robots to move through complex environments using only a single RGB camera, achieving 76.6% success on unseen R2R-CE benchmarks. This performance outperforms multi-sensor approaches while being more efficient. The model takes RGB images and plain-language instructions to guide robots through tasks such as navigating through a building or avoiding obstacles. Robostral Navigate uses a pointing-based navigation system combined with reinforcement learning for continuous improvement, setting the foundation for unified embodied AI in robotics. The model is capable of adapting to real-world obstacles unseen during training, making it versatile across various robot types and environments. It also operates efficiently on wheeled, legged, and flying robots, demonstrating robustness to differences in camera intrinsics and token-efficient training through prefix-caching. The release of Robostral Navigate marks a significant step in advancing embodied navigation for robotics applications.
Robostral Navigate is built entirely in-house and trained in simulation, with an efficient data generation pipeline that produced a dataset of approximately 400,000 trajectories across 6,000 scenes. The model's training process uses a tree-based attention-masking strategy to compress entire episodes into a single sequence, reducing training tokens by 22× while preserving learning signals. This approach transforms training runs that would take months into days. Additionally, the model leverages online reinforcement learning through CISPO, an algorithm that improves success rates by 3.2% after supervised training. Mistral emphasized that the model continues to show improvement with more training and experiments, highlighting its potential for further development.
Mistral stated that Robostral Navigate is the first step toward a unified embodied agent, with navigation as a foundational capability for general-purpose robotics. The model demonstrates that state-of-the-art embodied navigation can be achieved with a compact model and a single RGB camera. The release of Robostral Navigate opens new possibilities for applications in manufacturing, delivery, logistics, and hospitality, making it one of the most in-demand capabilities for customers. Mistral is also actively expanding its robotics team, seeking talented researchers and engineers to continue advancing embodied AI.
Source: mistral
