Skip to content
NewsModel release

Mistral releases Robostral Navigate, an 8B single-camera navigation model that beats depth-sensor systems on R2R-CE

· by Pondero Newsdesk

The short version

Mistral AI released Robostral Navigate on July 8, its first robotics navigation model. The 8B system scored 76.6% on R2R-CE for unseen environments using one RGB camera, 4.5 points above the leading multi-sensor approach and 9.7 points above the prior best single-camera system.

Mistral releases Robostral Navigate, an 8B single-camera navigation model that beats depth-sensor systems on R2R-CE

Mistral AI released Robostral Navigate on July 8, a navigation model that guides robots through unfamiliar environments using one standard RGB camera and no depth sensors. On the R2R-CE benchmark for previously unseen buildings, it scored 4.5 points above the best multi-sensor system per TestingCatalog.

What

The 8B model takes RGB images and plain-language instructions as inputs and outputs navigation commands for the robot, per Mistral's release page. It runs on wheeled, legged, and aerial platforms and is designed to generalize across different robot sizes and camera hardware. Robostral Navigate is Mistral's first announced robotics model; the company built it in-house without adapting an existing open-source base, per the announcement.

Training happened entirely in simulation: roughly 400,000 trajectories across 6,000 scenes, processed with a tree-based attention-masking strategy that cut training time from months to days per TestingCatalog. Mistral then applied online reinforcement learning to improve navigation scores.

On R2R-CE validation for previously unseen environments, the model scored 76.6% per TestingCatalog. That is 9.7 points above the prior best single-camera approach and 4.5 points above the leading multi-sensor system. Access is currently limited to select partners in manufacturing, logistics, delivery, and hospitality.

Why it matters

Depth sensors and LiDAR dominate commercial navigation deployments because camera-only approaches have historically underperformed them. A single-camera result that outscores those setups shifts the hardware calculus for operators managing mixed robot fleets or planning first deployments where sensor cost is a real constraint.

The 4.5-point margin against multi-sensor systems is the specific figure to stress-test in follow-on work. Benchmark performance in controlled simulation does not guarantee the same gap in real environments with variable lighting and crowded corridors. Results from Mistral's launch partners would be the first external signal on whether the margin survives production conditions.

What to watch next

Mistral has not published API pricing or a general-availability date for Robostral Navigate. The clearest external signal will come from launch partners sharing results outside benchmark conditions. A broader physical AI platform from Mistral, extending from navigation to manipulation or workflow control, could follow in the second half of 2026.

Sources