Reactive plan execution in multi-agent environments

Resumen

One of the challenges of robotics is to develop control systems capable of quickly obtaining intelligent, suitable responses for the regularly changing that take place in dynamic environments. This response should be offered at runtime with the aim of resume the plan execution whenever a failure occurs. The term reactive planning addresses all the mechanisms that, directly or indirectly, promote the resolution of failures during the plan execution. Reactive planning systems work under a continual planning and execution approach, i.e., interleaving planning and execution in dynamic environments. Most of the current research puts the focus on developing reactive planning system that works on single-agent scenarios to recover quickly plan failures, but, if this is not possible, we may require more complex multi-agent architectures where several agents may participate to solve the failures. Therefore, continual planning and execution systems have usually conceived solutions for individual agents. The complexity of establishing agent communications in dynamic and time-restricted environments has discouraged researchers from implementing multi-agent collaborative reactive solutions. In line with this research, this Ph.D. dissertation attempts to overcome this gap and presents a multi-agent reactive planning and execution model that keeps track of the execution of an agent to recover from incoming failures. Firstly, we propose an architecture that comprises a general reactive planning and execution model that endows a single-agent with monitoring and execution capabilities. The model also comprises a reactive planner module that provides the agent with fast responsiveness to recover from plan failures. Thus, the mission of an execution agent is to monitor, execute and repair a plan, if a failure occurs during the plan execution. The reactive planner builds on a time-bounded search process that seeks a recovery plan in a solution space that encodes potential fixes for a failure. The agent generates the search space at runtime with an iterative time-bounded construction that guarantees that a solution space will always be available for attending an immediate plan failure. Thus, the only operation that needs to be done when a failure occurs is to search over the solution space until a recovery path is found. We evaluated theperformance and reactiveness of our single-agent reactive planner by conducting two experiments. We have evaluated the reactiveness of the single-agent reactive planner when building solution spaces within a given time limit as well as the performance and quality of the found solutions when compared with two deliberative planning methods. Following the investigations for the single-agent scenario, our proposal is to extend the single model to a multi-agent context for collaborative repair where at least two agents participate in the final solution. The aim is to come up with a multi-agent reactive planning and execution model that ensures the continuous and uninterruptedly flow of the execution agents. The multi-agent reactive model provides a collaborative mechanism for repairing a task when an agent is not able to repair the failure by itself. It exploits the reactive planning capabilities of the agents at runtime to come up with a solution in which two agents participate together, thus preventing agents from having to resort to a deliberative solution. Throughout the thesis document, we motivate the application of the proposed model to the control of autonomous space vehicles in a Planetary Mars scenario. To evaluate our system, we designed different problem situations from three real-world planning domains. Finally, the document presents some conclusions and also outlines future research directions.