Robostral Navigate

What it is

Robostral Navigate is an embodied navigation model from Mistral AI, announced on 8 July 2026; it is the company's first model of this type^[2]. "Embodied" here means the model's output is an action rather than a sentence or a picture: it takes in RGB images and a plain-language instruction, and it moves a robot through the environment^[1].

The sensor requirement is deliberately modest. The model works from a single ordinary RGB camera, with no depth sensor and no LiDAR^[1]. A depth camera or a LiDAR unit gives a robot an explicit measurement of how far away things are; an RGB-only model has to infer that spatial structure from ordinary pictures instead. Dropping those sensors keeps the hardware side of a robot simpler, and it lets the model run on the kind of camera a robot already carries.

How it is built

Robostral Navigate has 8 billion parameters and was initialised from Mistral's in-house vision-language model^[1]. Starting from a system that already connects images with language gives the navigation model a base for reading a scene and understanding the instruction; training then teaches it to turn that reading into movement.

Mistral says the model was built entirely in house and does not rely on existing open-source vision-language models^[1]. Its data-generation pipeline was built entirely in simulation, and the model was trained entirely in simulation too^[1]. The training set consists of about 2.4 million simulated trajectories across 350,000 scenes^[1]. Simulated scenes can be generated and varied at a scale that would be slow and costly to capture in the physical world, which is the usual argument for this route: you can keep the training data growing without filming every corridor and room yourself.

Simulation training is followed by a reinforcement-learning stage. CISPO, an online reinforcement learning algorithm, is used to boost the model's performance after the simulated training^[1], and Mistral credits it with adding 3.2 points of success rate^[1]. The general idea of such a stage is that the model tries routes, is rewarded when it reaches the goal, and its parameters are nudged toward the behaviour that produced the reward.

How well it does

Mistral reports 76.6% success on the validation unseen split of R2R-CE and 79.4% on the validation seen split^[1]. On navigation benchmarks the "unseen" split holds environments the model was not trained on, which makes it the tougher of the two tests; the gap between the two numbers shows how much performance depends on whether the place has been seen before.

Where it sits

The same model is used on wheeled, legged and flying robots^[1]. Mistral also says it adapts to real-world obstacles that were not seen during training^[1].

The result is a navigation model built from a vision-language base, trained end to end in simulation, and paired with a reinforcement-learning step rather than with extra sensors. For a company whose public work has centred on language models, an embodied navigation model is a step from software that answers to software that acts in a physical space. It also illustrates a common pattern in robotics: generate the data and do the training in simulation, then rely on the trained model to carry over to machines in the real world.