Track-MDP++: RL for Target Tracking with General Controlled Sensing Models

State-of-the-art approaches for target tracking with controlled sensing (TTCS) are typically based on model-based Partially Observable MDP (POMDP) formulations. In this work, we address the model-free TTCS problem by proposing a novel MDP framework that accommodates arbitrary sensing models and exploring a Reinforcement Learning (RL) approach where the motion model for the target is unknown. We show that the infinite-horizon tracking reward of our model-free RL algorithm approaches that of the optimal POMDP policy that knows the object movement model. Simulations demonstrate the computational efficiency of our method and its superior resource utilization compared to baseline approaches, particularly as measured by Cost Per Track (CPT).