Mistral just released Robostral Navigate, an 8B parameter model built from scratch for robot navigation. It guides robots through complex environments using just one RGB camera — no depth sensors or multiple lenses needed. In tests on the R2R-CE benchmark, it hit a 79.4% success rate, beating both the best single-camera method and systems with extra hardware. The model was trained entirely in simulation on about 400,000 paths across 6,000 virtual spaces. It already works for wheeled, legged, and flying robots, though Mistral hasn’t revealed when or how it’ll be available. The company calls this the foundation for universal robotics and is already experimenting with reinforcement learning, which boosted success by 3.2 percentage points with no slowdown in sight. Why it matters: Mistral is moving beyond language into embodied AI, signaling that the next big competition in open-weight models might be about physical world tasks, not just chat.