Seeing Through Uncertainty: Free-Energy-Inspired Real-Time Adaptation for Robust Visual Navigation

arXiv cs.AI · 6 hours ago

Researchers introduced FEP-Nav, a framework for visual navigation that adapts in real-time to corrupted or degraded visual inputs by using the Free Energy Principle from neuroscience to minimize prediction errors. The method combines a top-down decoder with adaptive normalization to handle visual corruption without requiring gradient-based updates during inference. The framework restored navigation performance under visual corruption conditions and outperformed existing non-adaptive and adaptive baseline methods.