Plasticity Recovery for Active Vision Reinforcement Learning
Year: 2026
Type: conference-paper
Abstract: Deep reinforcement learning (DRL) is effective for control in 3D environments, but fixed viewpoints often lead to information loss due to occlusion, motion, and pose changes. Active vision, which allows agents to control sensors to select informative views, offers a promising solution by mimicking biological visual behaviors. However, existing DRL-based active vision methods suffer from unstable training, slow convergence, and challenges associated with reward shaping, often resulting in local optima or policy degradation. We identify plasticity loss-where network modules prematurely stabilize and fail to adapt to new experience-as a key cause. To address this, we propose the Inactivity Ratio Guided Plasticity Injection (IRGPI) method, which recovers plasticity in specific modules to enhance continual learning. Experiments across three active vision frameworks and four tasks demonstrate that IRGPI improves training stability and final performance.
Authors: Hanfei Zhao, Yan Zhang
Institution: Beijing Institute of Technology
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7224925255