Mistral released Robostral Navigate on [date], saying it enables robots to autonomously navigate complex environments using only a single RGB camera. It is the company's first model built for embodied navigation since its prior milestone in vision-language modeling.
Mistral reported 76.6% success rate on R2R-CE validation unseen, measured on the Room-to-Room in Continuous Environments benchmark. That compares with 79.4% success rate on validation seen.
Robostral Navigate is built on simulated data and token-efficient techniques and targets applications in manufacturing, delivery, logistics, and hospitality. Availability begins with a general release, initially for enterprise and research audiences.
"Leave the lobby, walk through the corridor, enter the supply room, and stop to face the second shelf," said Théo Cachet, AI Science Robotics Lead. The model completes the entire task on its own, moving through a live space full of people and obstacles it was never shown.
The announcement follows Mistral's expansion into embodied AI and robotics. The company emphasized that Robostral Navigate represents a significant step forward in autonomous navigation capabilities.
Mistral did not say how the model will handle navigation in highly dynamic environments, and it raised the open question of how to further improve success rates. The company said it is confident more training and experiments will continue to push the number up.
Source: mistral
