VAE+DDPG: An Attention‐Enhanced Variational Autoencoder for Deep Reinforcement Learning‐Based Autonomous Navigation in Low‐Light Environments

Depth sensors often lose range inside narrow corridors once ambient illumination approaches Lx, and perception planning deteriorates rapidly. Collisions with desks recur, although on semilogarithmic axes, the downturn appears trivial because lower luminance compresses ordinate values. The experiment grafts a variational autoencoder onto a feature pyramid attentive mechanism, then tunes the joint encoder alongside a Deep Deterministic Policy Gradient (DDPG) agent. This composite objective drifts the latent representation toward geometric features rather than raw intensity values. After convergence the learned policy achieves roughly % success at Lx and almost % when illumination reaches Lx inside cluttered Gazebo scenes during navigation trials. A plain DDPG stabilises near %, whereas freezing the encoder drags performance, confirming benefits of simultaneous tuning. Transfer trials inside an unseen maze reduce success by three points, implying that the latent space preserves obstacle silhouettes despite texture variation. Deployment processes depth frames at ten hertz on one RTX 4070, occupying GB memory. Control reverts to a reactive fallback once illumination falls below Lx in field experiments suddenly. Collectively, this evidence indicates that affordable indoor autonomy under considerable lighting variation appears attainable. The code can be found at https://github.com/uiseoklee/VAEplusDDPG .