Brain
Robostral Navigate is Mistral AI's first model for embodied navigation. An 8B parameter model that takes RGB images and natural language instructions to guide robots through complex environments. State-of-the-art on R2R-CE (76.6% success on validation unseen, 79.4% on validation seen).
Operates with a single RGB camera, no LiDAR or depth sensors. Built entirely in-house, trained in simulation on ~400K trajectories across 6,000 scenes. Generalizes across wheeled, legged, and flying robots.
Foundation model · Maturity: Research · Closed
Machine-readable surfaces
- Markdown mirror: /brains/robostral-navigate.md
- RSS feed: /brains/robostral-navigate/feed.xml
- JSON-LD: embedded in this page’s head
- REST API: /v1/brains/75e9dffc-773c-4ef7-b0a0-1de2a0f9a850
- Revision history: /brains/robostral-navigate/history
- Data documentation: /data
- Query this programmatically: Deploy MCP
Architecture
8B parameter VLM initialized from vision-language model for grounding (pointing, counting, object localization). Navigation via pointing (infers image coordinates of target location) with fallback to metric displacements. Trained with prefix-caching (tree-based attention masking, 22x token reduction) and online RL (CISPO algorithm, +3.2% improvement).
Key facts
- Type
- Foundation model (embodied navigation)
- Size
- 8B parameters
- Input
- RGB images + natural language instructions
- Sensors required
- Single RGB camera only; no LiDAR or depth sensors
- Benchmark
- State-of-the-art on R2R-CE: 76.6% success on validation unseen, 79.4% on validation seen
- Margin
- Beats best single-camera approach by 9.7 points; beats best multi-sensor by 4.5 points despite no depth sensors
- Training
- Built entirely in-house; trained in simulation on ~400K trajectories across 6,000 scenes; no reliance on existing open-source VLMs
- Efficiency
- Prefix-caching with tree-based attention masking compresses episode into single sequence; 22x token reduction vs per-timestep training
- RL post-training
- Online reinforcement learning (CISPO algorithm) improved success rate by 3.2% after supervised training
- Generalization
- Runs on wheeled, legged, and flying robots; generalizes across robot sizes and camera intrinsics
- Navigation method
- Pointing-based (infers target coordinates in camera view) with fallback to metric displacements when target out of view
- Authors
- Theo Cachet, Arjun Majumdar, Srijan Mishra, Thomas Chabal, Chris Bamford, Elliot Chane-Sane, Benjamin Tibi, Ludovic Ho Fuh, Olivier Duchenne (AI Science Robotics)
Developed by (1)
Common questions
What is Robostral Navigate?
Who developed Robostral Navigate?
Is Robostral Navigate open source?
What type of AI is Robostral Navigate?
What is Robostral Navigate's maturity stage?
Which robots run on Robostral Navigate?
Sources (3)
Methodology: Verified · 3 sources (no primary) · last reviewed 2026-07-14
Verification posture
Verified
Low confidence
Review state
Stable
Last reviewed 2026-07-14
Maturity + lifecycle
Maturity stage: research
Sources by quality tier
- 3
- unclassified
- Unclassified source
The framework is documented at /methodology. Corrections at /corrections. Reviewer: DEPLOY editorial team.
Methodology surface for Robostral Navigate.Canonical ID 75e9dffc-773c-4ef7-b0a0-1de2a0f9a850