A multi-agent reinforcement learning task allocation framework for navigation sequences of construction robots

Construction robotics, while steadily gaining research traction, encounter practical constraints limiting widespread adoption, with challenges stemming from the dynamic nature of construction environments and the inefficiencies inherent in single-agent robotic applications. Traditional programming methods prove insufficient for tasks subject to frequent changes, necessitating a new approach to task allocation and coordination for construction robots. Addressing these, our research introduces a transformative framework integrating Multi-Agent Reinforcement Learning (MARL) with construction robotics, advancing task allocation autonomy without prescriptive datasets or manual programming. Still, there are some big technical challenges. Most importantly, we need an environment to be modeled, benchmarked, and tested with various options. In order to build these training schemes. This thesis proposed a modular framework to implement multi-robot training: It allows robotic platforms, working environments, hardcoded construction activities, task logic, and training algorithms to be substituted to find the optimal solution in various construction scenarios. This research utilizes small-scale mobile robots to navigate various workspaces and perform diverse tasks, serving as a case study to validate the proposed framework. It specifically considers the mobile nature of construction robots as opposed to the static robots typically found in manufacturing, thereby addressing the dynamics, allowing for role adaptability, and enhancing the realism of simulations. Given the complexity and variability of the tasks, a 2D ROS-based simulation environment was developed, which has shown promising results in task scheduling and robot positioning among multiple construction robots. This framework paves the way for integrating robots with training algorithms across various construction settings, thereby boosting productivity and fostering wider acceptance of construction robotics.Despite its potential, the current framework has limitations. It is unclear which algorithms are most effective for training task allocation strategies on dynamic construction sites. However, the framework acts as a valuable testbed for further development and benchmarking of suitable algorithms. The model simplifies work logic and physical elements, leaving scope for addressing more complex scenarios, such as role changes and collaboration on construction sites. Moving forward, I plan to enhance the training environment to accommodate a broader range of scenarios. Users could adapt the navigation actions to other construction activities, such as manipulation and data collection, to get the optimal sequential orders. Future efforts will also focus on leveraging advanced simulators like ISAAC Nvidia for faster training and improved robot performance.