Mistral's Robostral Navigate steers robots with a single RGB camera, no LiDAR

Mistral has released Robostral Navigate, an 8-billion-parameter model that drives a robot through an environment from nothing more than a plain-language instruction and a feed from one ordinary RGB camera. Where competing systems lean on depth sensors, LiDAR, or arrays of cameras, this model uses none of them, yet scores 76.6% success on the R2R-CE ‘validation unseen’ benchmark—9.7 points ahead of the best single-camera method and 4.5 points ahead of the best depth- or multi-camera system. It runs on wheeled, legged, and flying robots and generalizes across robot sizes and camera intrinsics, with target applications in manufacturing, delivery, logistics, and hospitality.

The model is built entirely in-house rather than fine-tuned from an existing open-source VLM. It starts from Mistral’s grounding-focused vision-language model—one already good at pointing, counting, and object localization—and treats navigation as an extension of knowing where things are. Its core trick is ‘pointing’: predicting the image coordinates of where to move next plus the desired arrival orientation, which sidesteps sensitivity to world scale and camera calibration. When the target sits outside the current field of view, it falls back to metric displacements in the robot’s local frame. Training ran entirely on roughly 400,000 simulated trajectories across 6,000 scenes.

Two efficiency and quality levers stand out. A prefix-caching scheme with tree-based attention masking packs an entire episode into a single sequence, cutting training tokens 22× while preserving the learning signal and turning month-long runs into days. After supervised training, an online reinforcement-learning stage (CISPO) lets the model learn from failures and exploration, adding 3.2 points with no sign of plateauing. Mistral frames the model as a first step toward a unified embodied agent and is actively hiring for its robotics team.