Robustness of Vision Perception to Real-World Conditions

A mobile robot's object-detection model performs well in the lab (mAP ~0.9 on the validation set) but degrades badly when deployed in a warehouse: it misses objects under harsh overhead lighting, in motion blur from fast turns, and when objects are partially occluded by shelving. Walk through how you would diagnose the root causes of this performance gap and what concrete strategies you would apply to close it. Discuss data, model, and evaluation considerations.