robostral-navigate-d88cf620·2 events·first seen Aliases: Robostral Navigate
Mistral AI has released Robostral Navigate, a model targeting robotics navigation tasks and claiming state-of-the-art performance in that domain. This represents Mistral's entry into embodied AI and robotics, a significant expansion beyond their core language model work. The announcement is drawing notable community attention on Hacker News with 366 points and 87 comments.
Mistral AI has released Robostral Navigate, an 8B vision-language model enabling autonomous robot navigation using only a single RGB camera and no depth sensors or LiDAR. The model achieves 76.6% success rate on the R2R-CE validation unseen benchmark, outperforming the best single-camera approach by 9.7 points and multi-sensor systems by 4.5 points. Built entirely in-house with simulation-generated data (~400K trajectories across 6,000 scenes), it uses a novel prefix-caching training algorithm that reduces training tokens by 22× and online reinforcement learning via CISPO for further gains. The model generalizes across wheeled, legged, and flying robots and represents Mistral's entry into embodied AI.