Robust Dynamic Object Detection in Cluttered Indoor Scenes via Learned Spatiotemporal Cues

A team from MIT has published a new paper, "Robust Dynamic Object Detection in Cluttered Indoor Scenes via Learned Spatiotemporal Cues," presenting a significant leap forward for robot perception. The core challenge they address is the failure of standard LiDAR systems to detect moving objects—like a person walking near a bookshelf—when those objects are close to static structures or only partially visible. Current solutions that add cameras are limited by their field of view and inability to recognize novel objects. This new framework is LiDAR-only, making it more robust and hardware-simpler.

The system works by intelligently fusing two data streams. First, it uses a temporal occupancy grid to segment areas that are moving. Second, it employs a neural network trained to provide a bird's-eye-view "dynamic prior," essentially a learned prediction of where moving objects are likely to be. A fusion module uses this prior to recover detections that traditional geometric clustering would miss. In experiments with motion-capture ground truth, the method outperformed the state-of-the-art, achieving a 28.67% higher recall and an 18.50% higher F1 score in highly cluttered environments, while maintaining precision.

This advancement is crucial for the next generation of autonomous robots operating in dynamic human spaces like warehouses, hospitals, and homes. By significantly reducing false negatives—the dangerous scenario of a robot not seeing a moving obstacle—the technology paves the way for safer and more reliable cohabitation between humans and machines. The paper is available on arXiv under the identifier 2603.15826.