Selective Perception under Active Resource management for Reinforcement Learning

Abstract: Reinforcement learning in multi modal environments often relies on sensors that vary in informativeness, latency, and computational cost, making continuous full observation impractical under real resource constraints. Standard approaches typically embed sensing cost into the reward, treating all measurements as uniformly undesirable and failing to account for fixed sensing budgets that require informed allocation. We introduce Selective Perception under Active Resource management (SPAR), a framework that formulates sensor acquisition as a constrained decision problem in which an agent selectively activates feature groups while operating under a predefined budget. SPAR enforces this constraint through a Lagrangian relaxation that enables the agent to balance sensor acquisition costs with task performance dynamically throughout learning. To support stable optimization from partial observations, SPAR incorporates a dual-policy architecture that separates sensing from control, reducing the variance of combinatorial sensing gradients and preventing budget-driven penalties from disrupting reward-maximizing control action learning. Across multi modal benchmarks, SPAR achieves competitive and often superior performance compared to static feature subsets, random acquisition policies, and cost-penalized baselines, while satisfying budget constraints across varying resource levels. These results demonstrate that budget-aware selective perception offers a principled and effective foundation for reinforcement learning in resource limited settings.