Mistral AI Releases Robostral Navigate: 8B Single-Camera Embodied Navigation Model Outperforms Multi-Sensor Systems on R2R-CE

Mistral AI has introduced Robostral Navigate, an 8B parameter vision-language model that achieves state-of-the-art autonomous robotic navigation using only a single RGB camera. Trained on simulated data with a 22× token reduction via tree-based prefix-caching, the model outpaces multi-sensor architectures on unseen Room-to-Room in Continuous Environments (R2R-CE) benchmarks.

Architecture and Performance on Unseen Benchmarks

Mistral AI has launched Robostral Navigate, an 8B parameter model designed for autonomous embodied navigation using a minimal sensor footprint. Unlike traditional robotic navigation stacks that rely on sensor fusion across LiDAR, depth sensors, and multi-camera arrays, this system operates entirely on input from a single RGB camera alongside natural language instructions.

On the Room-to-Room in Continuous Environments (R2R-CE) unseen validation benchmark, the model achieves a 76.6% success rate. This outperforms the previous state-of-the-art single-camera approach by 9.7 points and surpasses the leading systems that utilize depth sensors or multiple cameras by 4.5 points. On seen validation environments, the model achieves a 79.4% success rate. The architecture generalizes across wheeled, legged, and aerial robotic platforms of varying form factors, maintaining robustness to variations in camera intrinsics.

Pointing-Based Target Inference and Fallbacks

To navigate without metric depth hardware, Robostral Navigate utilizes visual grounding to establish spatial relationships. The model is initialized from an in-house vision-language foundation model specialized for grounding tasks such as pointing, counting, and object localization.

The primary control loop relies on pointing-based navigation. Given a natural language instruction and observation history, the model infers the 2D image coordinates of the next target location within the current camera frame, alongside the target orientation upon arrival. By abstracting the control commands to image space coordinates rather than metric displacements, the policy remains robust to variations in camera intrinsics and physical robot scale. When a target falls outside the current camera field of view, the model switches to local coordinate metric displacements, generating relative vector instructions (such as translation and rotation offsets) in the robot's local frame.

Simulation Pipeline and Tree-Masked Prefix-Caching

Training an 8B model on sequential trajectory data presents significant compute bottlenecks. Mistral bypassed real-world telemetry collection by building a simulated data generation pipeline, gathering approximately 400,000 trajectories across 6,000 distinct virtual scenes.

To optimize training throughput, the supervised learning stage utilizes an attention-masking strategy with prefix-caching. This tree-based attention mechanism packs an entire navigation episode into a single sequence. Consequently, the model can compute gradients for all time steps within a single forward pass, while mathematically preventing forward-looking information leakage between temporal steps. This architecture choice reduces the total number of required training tokens by 22× compared to standard step-by-step sample generation, shrinking the training timeline from months to days.

Mitigating Covariate Shift via CISPO Reinforcement Learning

To address the compounding execution errors typical of pure behavior cloning, Mistral implemented post-training optimization. Following supervised learning, the model is refined using an online reinforcement learning algorithm named CISPO.

This online training loop allows the agent to experience environmental failures, run trial-and-error exploration, and learn recovery behaviors. By directly addressing the distribution shift of vanilla behavior cloning, CISPO improved the navigation success rate by 3.2% without showing signs of performance plateauing.