Mistral embodied navigation · released July 8, 2026 Single RGB camera · contact access

Quick answer

Robostral Navigate turns one camera view into a next move.

Mistral's 8B navigation model follows plain-language routes from RGB observations without depth or LiDAR in the published setup. The launch is a research and contact surface, not a public endpoint or a production-safety certification.

Verified provider record

Purpose

Embodied navigation from a natural-language instruction and a history of RGB observations.

Model size

Mistral describes an 8B model built in-house and trained entirely in simulation.

Sensor claim

One ordinary RGB camera, without depth sensors or LiDAR in Mistral's published setup.

Action representation

Points to a target location and desired orientation in the current image; falls back to local-frame displacement when the target is outside the field of view.

Published access

The launch routes readers to contact Mistral. It does not publish a self-serve API endpoint, downloadable weights, price, quota, or license.

Flowith access

Not established. A Mistral robotics announcement does not prove availability in Flowith.

Physical-system boundary

Benchmark the policy behind independent safety gates

Gate 1

Replay before motion

Test recorded and simulated observations first. Compare commanded route, predicted targets, recovery behavior, and stop conditions.

Gate 2

Isolate the course

Use a controlled environment with mapped hazards, no unconsenting bystanders, bounded speed, geofencing, and a physical emergency stop.

Gate 3

Add independent controls

Keep collision avoidance, localization checks, actuator limits, watchdogs, and safety-rated controls outside the model policy.

Gate 4

Escalate failures

Require a human supervisor and a safe-stop path for ambiguity, occlusion, camera loss, drift, blocked routes, and out-of-distribution scenes.

Provider-published evaluation

Published result

Mistral reports 79.4% success on R2R-CE validation seen and 76.6% on validation unseen.

Training record

Mistral says the simulation pipeline used about 2.4 million trajectories across 350,000 scenes.

Efficiency record

Mistral reports prefix-caching reduced training tokens by 22× while preserving its stated learning signals.

Post-training record

Mistral attributes a 3.2-point success-rate improvement to its online reinforcement-learning stage.

These are Mistral's launch results. Reproduce the relevant operating conditions and measure interventions, collisions, near misses, route completion, latency, drift, and safe-stop behavior on the intended platform.

Official source

  1. Mistral's Robostral Navigate release — architecture, sensor setup, pointing policy, benchmark results, training description, demonstrations, and contact path.
  2. Mistral contact — the access path linked by the launch for embodied-navigation work.

No public weights, API model card, pricing page, license, or Flowith selector record was available in the cited launch. Do not infer them from Leanstral or another Mistral model.

Continue by decision intent

Robot evaluation guide Translate the release into sensor, safety, privacy, test-course, and deployment gates. Leanstral 1.5 Use Mistral's separate formal-proof record when the task is verified Lean 4 work. Lean proof guide See a compiler-gated software workflow with different access and risk boundaries. More model guides Browse practical access, evaluation, and implementation decision records.

Robostral Navigate questions, answered

Robostral Navigate is Mistral's 8B embodied-navigation model for following natural-language route instructions from RGB camera observations.
Mistral's published system uses one ordinary RGB camera and no LiDAR or depth sensor. A production robot may still need independent sensing and safety controls for its operating environment.
The July 8 launch directs readers to contact Mistral and does not publish downloadable weights, a self-serve endpoint, API price, quota, or license. Verify those items directly with Mistral.
No. Mistral's R2R-CE results measure a defined navigation benchmark. They do not certify collision avoidance, functional safety, privacy, cybersecurity, reliability, or suitability for a specific robot and site.
Mistral shows and describes live-space navigation, but any real deployment needs a site-specific risk assessment, consent and camera controls, independent safety systems, bounded speed, emergency stop, supervision, and qualified engineering approval.
This page does not claim a Flowith mapping. Mistral's release and a Flowith model selector are separate sources, so verify the live workspace independently.