Mistral has introduced Robostral Navigate, its first AI model designed for robot navigation. The 8B model enables robots to navigate complex environments using only a single RGB camera, according to the company. It achieved a 79.4 percent success rate on the R2R-CE benchmark, a standard test for navigation in unknown environments. This performance surpasses both the best single-camera method and systems that use depth sensors or multiple cameras, according to Mistral.
The model was developed entirely in-house and trained in simulated environments, using about 400,000 recorded paths across 6,000 virtual spaces. It is compatible with wheeled, legged, and flying robots. Mistral has not yet disclosed availability details for the model. The company views navigation as foundational for universal robotics and plans to continue improving the model.
Mistral stated that experiments with reinforcement learning already increased the success rate by 3.2 percentage points, with no indication of a plateau. "We are confident that more training and experiments will continue to push this number up," the company said. Source: thedecoder
