r2r-ce-3719717a·1 events·first seen Aliases: R2R-CE
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.