Adaptive edge resource optimization for vehicle-edge cooperative perception with deep reinforcement learning

Abstract

Vehicle-base station cooperative perception extends sensing coverage and detection precision in intelligent transportation systems. However, random task arrivals, heterogeneous communication-computing resources and short vehicle coverage residence time bring great difficulties to efficient task scheduling. To solve this challenge, this paper constructs a deep reinforcement learning scheduling framework suitable for dynamic vehicular edge environments. We first establish a vehicle-base station cooperative perception model to describe perception tasks, vehicle mobility, channel states and heterogeneous computing resources. Based on compressed state representation, an optimized PPO algorithm is developed to jointly adjust task offloading schemes, transmission power and computing resource allocation. The designed policy adopts constraint-aware hybrid actions, invalid node masking, advantage normalization and entropy regularized clipping updates to stabilize training under hybrid discrete-continuous decision space. Furthermore, a reward function is created to balance task execution delay and response fluctuation. Simulation results verify that the proposed framework achieves lower average task latency, higher scheduling efficiency and more stable system performance compared with baseline algorithms.