RoboAtlas is a new active SLAM system that decides where a robot should look next by balancing the need to explore unknown space with the need to find semantically meaningful objects. It builds a 3D semantic map with a module called OpenRoboVox, then uses a contextual multi-armed bandit to choose among frontier-based exploration, global reasoning over the semantic map, and egocentric reasoning from a vision-language model. As the map improves, the system shifts from pure exploration to semantically guided navigation.

In real-world tests on a Unitree Go2 robot in environments larger than 1,800 square meters with roughly 30,000 mapped semantic instances, the system completed all tasks. On the GOAT-Bench "Val Unseen" benchmark, RoboAtlas achieved a 90.6% success rate when using GPT-4o, beating the strongest prior baseline by 17.8 percentage points.

Notably, swapping GPT-4o for the much smaller Qwen2.5-VL-7B still yielded an 88.8% success rate, which outperformed every baseline that used GPT-4o. The authors argue this shows that the information provided by the large-scale semantic map is more important than simply upgrading the foundation model. The results suggest that grounding foundation models in detailed 3D semantic maps is a promising path toward robust and efficient contextual active SLAM.