AmarZero: Vision-Based RL Agent for Object Tracking & Planning

AbstractThis paper introduces AmarZero, a novel model-based reinforcement learning agent designed for real-time object tracking and intelligent planning using advanced computer vision techniques. The proposed system integrates Vision Transformers (ViT) for high-level feature extraction, Kalman Filters for robust state estimation, model-based reinforcement learning for decision-making, and Monte Carlo Tree Search (MCTS) for optimal action planning. By learning a predictive world model, AmarZero can efficiently anticipate object trajectories, handle occlusion, and make real-time decisions in dynamic environments without requiring prior knowledge. The results show that AmarZero outperforms conventional tracking methods, achieving a significant boost in accuracy and robustness under various challenging scenarios. The implications of this work extend to autonomous navigation, robotic vision, and security applications.