Slow Brain, Fast Planner: Latency-Resilient VLM-Augmented Urban Navigation
# Trajectory Scoring Problem Limits Real-Time Robot Navigation
Autonomous sidewalk robots can quickly generate multiple movement options, but they often choose the wrong path despite better alternatives being available. Researchers identified this as a "trajectory scoring" failure—where the decision-making system ranks candidate routes incorrectly, causing robots to veer onto grass, move toward pedestrians, or head in unintended directions. The core issue: fast planning systems generate diverse options, but their evaluation methods struggle in complex urban environments.
The problem directly affects deployment reliability in mixed human-robot spaces. Autonomous delivery, security, and logistics robots operating in real-world sidewalk conditions depend on trajectory selection that handles unpredictable obstacles, pedestrian movement, and varied terrain. When scoring functions fail to identify optimal paths from available candidates, operators face increased intervention requirements and reduced autonomous range—forcing manual corrections that undermine system efficiency and safety assurances.
The research centers on bridging latency gaps between slow visual reasoning and fast planning execution. This suggests the current bottleneck isn't generating movement options, but accurately evaluating them under real-world timing constraints. Integration teams deploying autonomous systems in urban logistics will need to evaluate whether their platforms' trajectory evaluation methods account for this scoring-accuracy gap alongside their path-generation speeds.