Robostral Navigate

Robostral Navigate is an 8-billion-parameter vision language model developed by Mistral AI for instruction-following robot navigation. Released on July 8, 2026, it receives a natural-language instruction and a history of images from one monocular RGB camera, then predicts waypoints for moving through the environment. Mistral introduced it as the company's first model for embodied AI navigation, while its model registry classifies it as an active specialized base model.^[1][2][3]

The 8B model is the planning component of a larger control pipeline. Its waypoint predictions feed a separate 121-million-parameter diffusion policy, followed by a platform-specific motion controller. Mistral's legal documentation describes the model as specialized solely for navigation, rather than manipulation or general-purpose AI.^[2][3]

Release and scope

Mistral announced Robostral Navigate on July 8, 2026. A technical paper by the model's contributors followed on July 22. The announcement and paper use the same product name, parameter count, input modality, and navigation-only task definition.^[1][3]

Scroll sideways for more →

Mistral did not provide a public checkpoint, model license, API identifier, price, code repository, or deployment guide in the announcement or paper. The CC BY 4.0 license shown by arXiv applies to the paper, while the cited materials do not assign that license to model weights or training data. Mistral's governance page says an EU general-purpose-model technical package is not required because Robostral Navigate is a specialized navigation model.^[1][2][3]

Architecture and outputs

Robostral Navigate was initialized from an internal dense 8B model trained for spatial grounding tasks such as pointing, counting, and object localization. A vision encoder converts each frame to visual tokens, which are appended to the tokenized instruction. The image history lets the model track earlier locations and progress through a route. Mistral does not identify the grounding model or disclose the vision encoder, layer counts, hidden dimensions, or attention configuration.^[1][3]

The deployed navigation stack has three stages:^[3]

Scroll sideways for more →

The preferred waypoint format uses pixel coordinates (u, v) for the farthest visible point on the intended route, together with a desired change in yaw. During training, the model also predicts local translation and rotation (delta x, delta y, delta theta). If no future waypoint is visible, as when the robot must turn around, it omits the pixel coordinates and uses the local displacement instead. About 10% of the training examples have an invisible destination. This split between a visual waypoint planner and later control stages differs from a monolithic vision-language-action model that emits low-level robot actions directly.^[3]

Pointing in image space reduces dependence on fixed camera intrinsics and robot scale, but physical integration is still specific to each platform. The diffusion policy needs the robot's height and radius, and the last controller in the stack is replaced for each embodiment.^[3]

Training

Mistral trained the system entirely in simulation. Its data-generation pipeline produced about 2.4 million expert navigation trajectories across 350,000 indoor and outdoor scenes. Each sample pairs a language instruction with RGB observations and corresponding navigation actions. The authors used farthest-point sampling to vary start and goal locations, including some routes that span multiple floors, and varied layout, object density, lighting, and architectural style.^[1][3]

Robot and camera properties were randomized for both the VLM and diffusion-policy data. Simulated robot height ranged from 0.4 to 1.8 m, radius from 0.15 to 0.45 m, camera placement from 70% to 100% of robot height, and camera pitch from 0 to 25 degrees. These ranges were intended to reduce dependence on one body shape or camera mount.^[3]

For supervised training, an entire episode is packed into one sequence of interleaved observations and actions. A prefix-tree attention mask lets each prediction use the shared instruction and observation history without seeing ground-truth actions from other time steps. The authors report that this changes token growth with episode length from quadratic to linear, reduces training tokens by 22-fold, and cuts runs that would take months to days.^[1][3]

The supervised model was then refined with online reinforcement learning using CISPO. The authors selected 35,000 difficult tasks on which the supervised policy did not reliably reach the goal. The terminal reward depends on geodesic distance and becomes flat within 2 m, encouraging the model to emit STOP instead of maneuvering indefinitely near the destination. Training required concurrent GPU workloads for simulation and rendering, vLLM-based action generation, and distributed updates, but the paper gives no accelerator type, GPU count, training hours, energy use, or cost.^[3]

Evaluation

R2R-CE is the continuous-environment form of the Room-to-Room vision-and-language navigation task. It removes the fixed navigation graph and perfect-localization assumptions used in earlier versions of the task. RxR extends instruction-following navigation with longer routes and multilingual instructions; Robostral Navigate was evaluated on the English-only RxR-CE split.^[4][5]

The paper defines success as stopping within 3 m of the goal. Oracle success records whether the agent came within 3 m at any point, navigation error measures final geodesic distance, and success weighted by path length (SPL) penalizes detours. The following figures are author-reported validation results:^[3]

Scroll sideways for more →

The later paper reports a higher R2R-CE unseen result than the release post, 77.4% rather than 76.6%. It also reports a 4.03-point gain from online reinforcement learning, from a 73.40% supervised baseline to 77.43%, whereas the announcement described a 3.2-point gain. Neither source explains the difference between the two reports.^[1][3]

In the paper's R2R-CE table, the strongest previous single-camera result was 66.9% from Qwen-RobotNav-4B. A Qwen-RobotNav-8B configuration using depth or multiple cameras reached 72.1%. Robostral Navigate's 77.4% result was therefore 10.5 percentage points above the former and 5.3 points above the latter. On RxR-CE, it exceeded the single-camera Qwen-RobotNav-8B result in success and SPL, but the depth-assisted Qwen configuration had a higher success rate, 76.5% versus 75.1%.^[3][6]

Robot demonstrations and runtime

The paper documents deployment on two physical platforms, the Galaxea R1 and Hiwonder JetAuto. Both used the same Robostral Navigate and diffusion-policy weights; only the low-level controller differed. Mistral's announcement also includes an autonomous long-horizon route through a working office and says the system can run on wheeled, legged, and flying robots. The paper's named cross-platform demonstrations, however, are limited to the R1 and JetAuto.^[1][3]

The published rates of 0.5 Hz for waypoint planning, 10 Hz for trajectory generation, and 100 Hz for motor control describe the intended pipeline timing. The sources do not state the inference hardware, memory requirement, power draw, end-to-end latency, maximum safe speed, or whether those rates were achieved on each demonstrated robot.^[3]

Safety and limitations

The headline benchmark results do not evaluate the complete physical control stack. For R2R-CE and RxR-CE, the authors used an AI Habitat pathfinder to move between the model's predicted waypoints. The results therefore measure simulated instruction following and waypoint planning, not the learned diffusion policy, embodiment-specific controller, or actuator behavior on a physical robot.^[3][4][5]

The release post and paper do not report collision rates, near misses, human interventions, camera-obstruction tests, adversarial instructions, emergency-stop behavior, or other physical-safety evaluations. The office route and two-platform deployment are qualitative demonstrations, with no stated number of trials or aggregate real-world success rate. They do not establish safe unattended operation around people.^[1][3]

A single RGB camera also creates a field-of-view limitation. The model cannot point to a waypoint it cannot see and must switch to a metric displacement prediction. VLM inference introduces a delay between the frame used for waypoint planning and the robot's current frame; the diffusion policy receives both frames, but no worst-case delay or high-speed stress test is disclosed.^[3]

Training solely in simulation avoids physical data collection but leaves real-world performance dependent on the simulated scene distribution and the chosen body and camera randomization ranges. The sources show transfer to two named platforms but provide no quantitative sim-to-real study. They also provide no public weights, data, or evaluation code for independent reproduction.^[1][3]

References

1. Mistral AI. Introducing Robostral Navigate. July 8, 2026. ↩
2. Mistral AI Legal Center. Robostral Navigate model lifecycle and classification. 2026. ↩
3. Majumdar, Arjun; Sooriyarachchi, Avinash; Tibi, Benjamin; et al. Robostral Navigate. arXiv:2607.20785v1, July 22, 2026. ↩
4. Krantz, Jacob; Wijmans, Erik; Majumdar, Arjun; Batra, Dhruv; Lee, Stefan. Beyond the Nav-Graph: Vision-and-Language Navigation in Continuous Environments. ECCV 2020. ↩
5. Ku, Alexander; Anderson, Peter; Patel, Roma; Ie, Eugene; Baldridge, Jason. Room-Across-Room: Multilingual Vision-and-Language Navigation with Dense Spatiotemporal Grounding. EMNLP 2020. ↩
6. Zhang, Jiazhao; Zhou, Gengze; Yin, Hale; et al. Qwen-RobotNav Technical Report: A Scalable Navigation Model Designed for an Agentic Navigation System. arXiv:2606.18112, June 2026. ↩