Definition
Robot navigation is the capability for mobile robots to follow natural-language instructions through environments, often evaluated on benchmarks such as R2R-CE (Room-to-Room Continuous Environments). Modern systems increasingly use vision-language models with single RGB cameras rather than LiDAR/depth stacks.
Key Points
- robostral-navigate: 8B single-RGB model; 76.6% R2R-CE unseen (Mistral-reported)
- Pointing-based navigation + RL (e.g., CISPO) improves robustness
- Sensor minimalism reduces hardware cost for logistics/hospitality robots