Efficient Semantic Navigation for Autonomous Robots

Summary & Methodology Analysis

OptiSight improves navigation efficiency by separating high-level semantic reasoning from low-level geometric movement. The system utilizes Grounded-SAM to identify target objects and extract their semantic boundaries in the image plane, which are then projected into a 3D camera reference frame. By handling geometric control separately, the framework limits the need for continuous Vision-Language Model (VLM) calls, which are computationally expensive. Instead, reasoning is triggered only at critical state transitions to optimize resource consumption. The framework was evaluated in AI Habitat across various indoor scenarios, demonstrating its ability to perform navigation without prior task-specific training. The initial twelve scenarios utilized the Qwen3.5-2B model for VLM reasoning, while the subsequent twelve experiments employed the Moondream2-2B model. Success was measured based on a defined mission-success criterion that tracks task completion and collision-free movement. Despite the success in multiple test cases, the framework remains sensitive to environmental constraints. The authors report that the system currently experiences collisions when navigating around closely positioned obstacles, indicating that maintaining sufficient clearance is a persistent challenge. Navigation failures in experiments 13, 16, and 17 further suggest that the system becomes less reliable when forced to maneuver through constrained passages.