Mistral's Robostral Navigate: a state of the art robotics navigation model
Points and comments are a snapshot, not live.
Mistral's Robostral Navigate achieves 76.6% on R2R-CE using only a single RGB camera.
Robostral Navigate is an 8B model for embodied navigation that takes RGB images and natural-language instructions to move robots through environments. It uses pointing (predicting image coordinates of the target) and local displacement fallbacks. Trained entirely in simulation on ~400,000 trajectories across 6,000 scenes, it uses prefix-caching to reduce training tokens by 22x. Online reinforcement learning via CISPO boosted success rate by 3.2%. The model runs on wheeled, legged, and flying robots and generalizes across robot sizes and camera intrinsics.
What commenters are saying
Commenters split on the model's practical value. Some praised Mistral's niche strategy as smart for competing with larger labs, noting applications in industrial automation. Others argued 80% success rate is practically useless in robotics-20% failure is unacceptable for real-world deployment. Skeptics compared it to early autonomous driving demos that proved harder than expected. Supporters countered that pointing-based navigation is a strong design choice and that smaller models offer latency advantages for local robot inference. Several flagged military applications already exist in Ukraine.