Residual-gated differentiable plasticity as an alternative to meta-learning in Target-Driven Visual Navigation
Matheus Santos Araújo, Yuri Lenon Barbosa Nogueira, Creto Augusto Vidal, Joaquim Bento Cavalcante-Neto · Computers & Graphics · 2026
Target-Driven Visual Navigation requires an embodied agent to reach a target object using egocentric visual input and a semantic label. In this work, we explore a complementary alternative formulation in which fast adaptation can be supported by differentiable plasticity rather than explicit meta-learning. We introduce an end-to-end reinforcement learning framework that integrates visual semantics and adaptive memory through a Residually-Gated Differentiable Plastic Hebbian Layer (RDPHL), allowing in-episode adaptation without explicit inner-loop optimization at train time. The architecture produces complementary attention representations, including a navigation-aware attention map (navCAM) and a Hebbian-based attention map, which help align perception and decision-making during navigation. Experiments in AI2-THOR show that the method achieves competitive overall performance and improves challenging-path navigation by 2.0 percentage points ( ≈ +6% relative improvement) in success compared to meta-learning and transformer-based baselines. These results suggest that differentiable plasticity is a lightweight alternative for online adaptation in embodied visual navigation, particularly in challenging-path scenarios.