Mistral's first robotics model beats multi-sensor systems with one RGB camera

Mistral released Robostral Navigate, an 8B model that steers wheeled, legged, and flying robots through unseen spaces from a single RGB camera — no LiDAR or depth — scoring 76.6% on the R2R-CE unseen benchmark, 4.5 points above the best multi-sensor system. The signal is that grounding-first VLM priors plus simulation-only training let a compact camera-only model match sensor-heavy stacks, pushing capable embodied navigation toward low-cost robots in logistics, delivery, and hospitality.

Source: mistral.ai ↗