- Robostral Navigate is an 8B model for autonomous robot navigation using only a single RGB camera.
- It achieves 76.6% success on unseen R2R-CE benchmarks, outperforming multi-sensor approaches.
- The model combines pointing-based navigation with reinforcement learning for continuous improvement.
- Built entirely in-house with simulation-trained data and token-efficient training techniques.
- It generalizes across robot types (wheeled, legged, flying) and adapts to real-world obstacles.
- Efficient training via prefix-caching reduces token use by 22x, speeding up training significantly.
- Online reinforcement learning (CISPO) boosts performance by 3.2% and enables recovery from failures.
- The technology unlocks applications in manufacturing, delivery, logistics, and hospitality.
- Robostral Navigate is a step toward unified embodied AI, paving the way for general-purpose robotics.