French AI company Mistral AI has released Robostral Navigate, its first robotics model, as part of an expansion into physical AI for industrial automation. The 8B parameter model enables robots to navigate using only a single camera, achieving a 79.4% success rate on a key benchmark. Trained entirely in simulation, the hardware-agnostic model is designed for a wide range of robots in factories and warehouses.
Robostral Navigate model
- ▪On July 8, 2026, Paris-based Mistral AI unveiled its first robotics model, Robostral Navigate
- ▪Robostral Navigate is an 8B parameter model that enables robots to navigate complex environments using a single camera and language prompts
- ▪The model is hardware-agnostic and designed to work with various suppliers' wheeled, legged, and flying robots
- ▪The system does not require lidar, advanced sensors, or multiple-camera setups for navigation
Technical performance benchmarks
- ▪Robostral Navigate achieves a success rate of up to 79.4% on the R2R-CE benchmark for navigation in unknown environments
- ▪Experiments using reinforcement learning have already boosted the model's success rate by 3.2 percentage points
- ▪Mistral AI states the model's performance surpasses both the best single-camera methods and systems using depth sensors or multiple cameras
Training methodology
- ▪The training data consisted of about 400,000 recorded paths across 6,000 different virtual spaces
- ▪The model was built entirely in-house by Mistral AI and trained exclusively in simulated environments
Physical AI expansion strategy
- ▪Mistral AI has not yet released details on the availability of Robostral Navigate
- ▪The release follows Mistral AI's acquisition of Austrian company Emmi AI in May
- ▪The launch is part of Mistral AI's expansion into physical AI for factories, warehouses, and industrial automation
Story comments
Loading comments…