Mistral AI introduced Robostral Navigate, an 8B parameter model designed for autonomous robot navigation using only a single RGB camera. The model achieves 76.6% success on unseen R2R-CE benchmarks, outperforming multi-sensor systems despite its simpler input. This performance leap is attributed to a combination of in-house simulation training, token-efficient training techniques like prefix-caching, and online reinforcement learning using the CISPO algorithm.

Robostral Navigate employs a pointing-based navigation strategy, inferring target coordinates directly from camera views, which enhances robustness to scale and camera variations. When pointing is not feasible, it falls back to local coordinate displacements. The model was trained on approximately 400,000 trajectories across 6,000 simulated scenes, with its training process optimized to reduce token usage by 22x while preserving learning signals.

This development holds significant implications for embodied AI, enabling applications in manufacturing, delivery, logistics, and hospitality. The model's ability to generalize across different robot types—wheeled, legged, and flying—and adapt to real-world obstacles unseen during training marks a step toward unified embodied agents for general-purpose robotics.