Mistral ships Robostral Navigate, an 8B single-camera robotics navigation model

Category: AI
Status: developing
Sources: Mistral writeup, HN discussion
Summary: Mistral published Robostral Navigate, an 8B vision-language model for robotic navigation that takes RGB camera frames and a natural-language instruction and moves a robot through an environment using a single camera. Mistral says it is initialized from a vision-language grounding model and navigates by pointing, predicting the image coordinates of the next target. Reported training uses about 400,000 simulated trajectories across 6,000 scenes, prefix-caching that cuts tokens 22 times, tree-based attention masking, and online reinforcement learning with the CISPO algorithm. On the R2R-CE benchmark Mistral reports a 79.4 percent success rate on validation-seen and 76.6 percent on validation-unseen, which it states beats the best single-camera approach by 9.7 points and the best multi-sensor systems by 4.5 points. The post does not state a license or weight availability.
Why it matters: A frontier lab extending the vision-language-action line into single-camera navigation is a concrete robotics data point, though the benchmark figures are vendor-reported and unreproduced.