Mistral AI just dropped Robostral Navigate, an 8-billion-parameter model designed to help robots find their way around using nothing but a single RGB camera. Give it a command like "leave the lobby, go down the hall, and enter the storage room," and the robot figures out the rest from its camera feed.
The real trick? It barely needs any hardware. No lidar, no depth sensors, no multi-camera rigs. Just one ordinary camera. In the R2R-CE benchmark, Robostral Navigate hit a 76.6% success rate in unfamiliar environments — that's 9.7 percentage points better than the best single-camera system and 4.5 points better than the top multi-sensor setup.
The model was trained entirely on simulated data, with about 400,000 paths across 6,000 scenes. It first looks at the camera feed to decide its next move; if the target isn't in view, it switches to step-by-step action commands like "move forward 2 meters, turn left 25 degrees." The team also reused cached computation from repeated content to slash training token usage by 22x, and reinforcement learning added another 3.2% to the success rate.