Robostral Navigate
Robostral Navigate is Mistral’s announced model for embodied navigation. It combines camera observations with a written instruction and was trained entirely in simulation.
Access not confirmed
Pointing and displacement outputs
The model normally predicts a location in the current camera view and an arrival orientation. For targets outside the view, it can instead produce displacement instructions in the robot’s coordinate frame. The release uses one RGB camera without depth sensors or LiDAR.Robostral Navigate release announcement
Research publication does not establish model access
The announcement reports an 8B model trained with about 2.4 million trajectories across 350,000 scenes. It publishes research results, but no public weight download or callable API is established in the checked release and lifecycle records. Access is therefore unconfirmed.Robostral Navigate release announcementRobostral Navigate primary source
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Access not confirmed |
| Launch date | |
| Weight license | No public weight license established by the announcement or lifecycle record.Robostral Navigate primary source |