SLAM-Assisted Transformer-Based TD3 Deep Reinforcement Learning for Adaptive Navigation and Mapping and Dynamic Obstacle Avoidance in Autonomous Construction Robotics

This research introduces a SLAM (simultaneous localization and mapping)-assisted deep reinforcement learning (DRL) framework that integrates a transformer-based twin delayed deep deterministic policy gradient (TD3) for efficient navigation and obstacle avoidance in differential drive robots. A real-time SLAM is integrated to provide accurate localization, real-time occupancy grids, and precise robot pose estimates, which are integrated into the TD3 reward estimation to enhance navigation precision. Key advancements include the integration of transformer-based architecture within TD3 to enhance generalization and robustness in dynamic construction environments. The system is trained in a simulated environment using ROS2 and Gazebo with domain randomization to introduce variations in lighting, sensor noise, and dynamics, thus bridging the simulation-to-reality gap (Sim2Real). Key findings include a 90% task success rate in navigating complex construction site layouts with dynamic obstacles. The enhanced TD3 model demonstrated significant adaptability by navigating unpredictable layouts and avoiding continuously changing obstacles while maintaining operational efficiency.