RoboAtlas: Contextual Active SLAM

How can a mobile robot effectively balance the need for broad geometric exploration with the need for targeted, semantically informed navigation when performing tasks in unknown environments? The authors aim to bridge the gap between traditional Active SLAM, which focuses on map completeness, and modern vision-language-driven navigation, which focuses on high-level semantic goals.

The authors introduce RoboAtlas, a framework that treats navigation as a dynamic goal-selection problem. It employs a mixture-of-experts approach where three distinct modules propose navigation targets: a frontier-based expert for geometric coverage, a semantic-map expert that reasons over a global 3D map, and an egocentric VLM expert that processes immediate visual observations. A contextual multi-armed bandit algorithm manages these experts, learning to prioritize exploration when the environment is unknown and shifting toward semantic search as the robot accumulates map data and contextual evidence. The system is supported by OpenRoboVox, a custom, memory-efficient 3D semantic mapping framework designed for real-time deployment on edge robotic platforms.

Key Findings

RoboAtlas demonstrates significant improvements in task success rates compared to existing baselines. In large-scale real-world environments exceeding 1800 square meters, the system achieved a 100% task success rate. On the GOAT-Bench "Val Unseen" benchmark, RoboAtlas achieved a 90.6% success rate using GPT-4o, outperforming the strongest prior baseline by 17.8 percentage points. Notably, when using a much smaller, computationally efficient model (Qwen2.5-VL-7B), the system still achieved an 88.8% success rate, suggesting that the framework's semantic mapping and adaptive decision-making are as critical to performance as the underlying foundation model itself.

This work provides a scalable, real-time solution for robots to perform complex, language-conditioned tasks in large, previously unseen spaces. By integrating semantic reasoning directly into the SLAM pipeline, RoboAtlas enables robots to move beyond simple coverage-based exploration, allowing them to act as intelligent agents that understand the spatial and semantic structure of their surroundings.