France's flagship AI lab is moving into robotics. Mistral AI has unveiled Robostral Navigate, an 8-billion-parameter vision-language model that navigates robots using only a stream of monocular RGB images — no depth sensors, LiDAR, multi-camera rigs or pre-built maps. The paper, published on arXiv alongside a Mistral research post, reports new state-of-the-art results on the two standard vision-and-language navigation benchmarks in continuous environments.
Pointing Instead of Metric Coordinates
Robostral Navigate decomposes navigation into two problems. The 8B VLM handles instruction understanding, scene reasoning and coarse path planning, predicting the next waypoint by pointing to image coordinates of the target location in the current camera view, together with the desired heading on arrival. Because waypoints live in image space rather than robot-specific metric coordinates, the policy is naturally robust to changes in camera intrinsics and scene scale. When the destination lies outside the field of view, the model falls back to predicting metric displacements, and it learns to output STOP on completion.
A lightweight 121M-parameter diffusion policy then converts each waypoint into a dense one-second trajectory at 10 Hz, and an embodiment-specific motion-tracking controller turns that into 100 Hz actuator commands. The VLM itself runs at 0.5 Hz — a slow-thinking, fast-acting split that mirrors the dual-system designs gaining traction across robot foundation models.
Simulation-Only Training, Compressed 22×
The team trained entirely in simulation, generating 2.4 million trajectories across 350,000 scenes — eliminating real-world data collection. A prefix-caching recipe packs entire episodes into single training sequences, cutting training tokens by 22× and compressing month-long training runs into days, while a tree-based attention mask prevents the model from conditioning on previous ground-truth actions. A final round of online reinforcement learning (CISPO) on a curated hard subset adds another 4 points of success rate.
Results: Beating Depth-Based Systems With One Camera
On R2R-CE (validation unseen), Robostral Navigate reaches a 77.4% success rate — 10.5 points above the best monocular method (Qwen-RobotNav-4B, 66.9%) and 5.3 points above the strongest depth- or multi-camera system (Qwen-RobotNav-8B, 72.1%). On RxR-CE, it sets another state of the art at 75.1% success and 68.7% SPL, outperforming all single-camera baselines with better path efficiency than depth-assisted models.
Same Weights, Different Robots
To demonstrate cross-embodiment transfer, the team deployed the identical VLM and diffusion weights on two very different platforms — the Galaxea R1 and the Hiwonder JetAuto — randomizing robot height (0.4–1.8 m), radius, camera placement and pitch during data generation. Only the low-level platform controller changed between robots.
The strategic signal matters as much as the benchmark numbers: Europe's most valuable AI lab is now publishing embodied AI research with a scalability-first recipe — minimal sensor assumptions, simulation-only training, embodiment-agnostic deployment. If the approach holds up beyond navigation, it sketches a low-cost path to general robot policies that do not depend on expensive sensing stacks.
