ai-trend-notifier
← wiki

$ cat wiki/models/robostral-navigate.md

Robostral Navigate

Spec

AttributeValue
DeveloperMistral AI
Released2026-07-08
Announced2026-07-08
Context windowunknown
Pricingunknown
LicenseTBD (not open-weight in initial release)
Availabilityunknown
Parameters8B
Sensorssingle RGB camera (no depth sensors, no LiDAR required)
Inputcurrent camera image + natural-language navigation instruction
Outputpointing — image coordinates of the target location + desired orientation on arrival
Training environmentsimulation only (no real-robot data required)
RL algorithmCISPO (online RL; +3.2 percentage points success-rate improvement)
Deploymenthardware-agnostic (any robot platform)

Release Date

2026-07-08. Announced via Mistral blog (source) and covered by Bloomberg.

Benchmarks

BenchmarkRobostral NavigatePrior best (single-camera)Prior best (depth/multi-cam)
R2R-CE (Room-to-Room in Continuous Environments) — validation unseen76.6%~66.9%~72.1%
Outperforms all prior navigation systems including those using depth sensors or multiple cameras — despite using neither.

Method

Robostral Navigate predicts robot movement via pointing: given the current camera frame, it outputs the pixel coordinates of the target destination and the desired approach orientation. This is robust to changes in camera intrinsics and world scale, unlike metric-displacement command approaches. Training is entirely simulation-based, using the CISPO online RL algorithm for policy improvement.

Use Cases

Autonomous robot navigation in complex physical environments:

  • Manufacturing (assembly line navigation, floor robots)
  • Delivery (last-mile indoor/warehouse robots)
  • Logistics (warehouse navigation)
  • Hospitality (front-of-house service robots)

Significance

Robostral Navigate is Mistral's first robotics model — a significant expansion from the language and coding domain into physical AI. Key points:

  1. Sim-only training closes the sim-to-real gap: achieving state-of-the-art R2R-CE with purely simulation training suggests the field's sim-to-real transfer problem is closer to solved than previously believed.
  2. Single-camera constraint as a feature: most deployed robots already have RGB cameras but not depth or LiDAR. A SOTA model requiring only a camera dramatically lowers the hardware bar for adoption.
  3. 8B class: small enough to run on-device (robot edge compute), unlike 7B+ VLM-based approaches requiring cloud inference.
  4. Hardware-agnostic: the pointing-based interface separates perception from actuation, making the model portable across robot hardware.

Compared To

ModelOrgSensorsR2R-CENotes
Robostral NavigateMistralRGB only76.6%New SOTA; sim-only training
Gemini Robotics ER 1.6Google DeepMindMulti-sensorIndustrial inspection (gauge-reading 93%); different task
Cosmos 3 SuperNVIDIAMulti-modalWorld model / generalist; different approach

Sources

Related

Referenced by

Sources