Robostral Navigate
8B embodied navigation model for autonomous robot movement using single RGB camera and natural language instructions
Context window
Not announced
Input / 1M tokens
Not announced
Output / 1M tokens
Not announced
Provider
Mistral AI
Data verified 2026-07-15
Robostral Navigate is an 8-billion-parameter vision-language-action model for robotic navigation that enables robots to autonomously navigate complex environments using only a single RGB camera and plain-language instructions. It achieves 76.6% success on the R2R-CE benchmark without requiring LiDAR, depth sensors, or multiple cameras. The model generalizes across wheeled, legged, and flying robot platforms.
Capability index
Relative estimates (0-100) to place this model against its peers, grounded in published benchmarks.
How to access it
Currently available to select partners in manufacturing, logistics, delivery, and hospitality sectors, with plans for broader access as testing continues.
Strengths
- ✓Single-camera design reduces hardware costs and complexity
- ✓High performance on R2R-CE benchmark (76.6% unseen, 79.4% seen)
- ✓Generalizes across different robot types and sizes
- ✓Trained entirely in simulation for rapid iteration
- ✓Compact 8B parameters suitable for edge inference
Best for developers who...
When to choose it (and when not to)
Reach for Robostral Navigate when...
- →When deploying navigation systems with minimal sensor hardware requirements
- →For cost-sensitive robotics applications in controlled indoor environments
- →When hardware-agnostic deployment across multiple robot types is needed
Look elsewhere if...
- ✕Safety-critical outdoor autonomous systems requiring depth and range sensing
- ✕Environments with challenging visual conditions (glass, mirrors, low-light areas)
- ✕Applications where latency is critical (edge inference numbers not published)
How to use it
- ›Use clear, sequential language instructions describing the desired navigation path
- ›Break complex routes into logical waypoints in plain language
- ›Describe target locations relative to visible features in the current field of view
Quickstart
Python# Pseudo-code example from official documentation
observation = camera.read() # RGB frame
instruction = "Leave the lobby, walk through the corridor, enter the supply room"
action = model.predict(instruction, observation_history)
if action.type == "point":
robot.move_to_pixel(action.x, action.y, action.heading)
else:
robot.move_local(action.forward_m, action.left_m, action.turn_deg)Integration requires mapping model outputs to robot velocity controllers; Mistral provides reference implementations for ROS and proprietary stacks
API model id: robostral-navigate
Benchmarks
| Benchmark | Score | Notes |
|---|---|---|
| R2R-CE (Room-to-Room in Continuous Environments) - Validation Unseen | 76.6% | Outperforms best single-camera approach by 9.7 points and best multi-sensor systems by 4.5 points |
| R2R-CE (Room-to-Room in Continuous Environments) - Validation Seen | 79.4% | Performance on previously seen environments |
Source: Mistral AI Official Announcement
Compare Robostral Navigate
Compare Robostral Navigate with any other model
Build a comparison →All model comparisons →