Work

Design of a Framework for Multi-Agent Team Cooperation using Behavior Trees

Year: 2026
Type: dissertation
Abstract: This thesis presents the development, implementation, and validation of a distributed control architecture designed to support the cooperation and coordination of multi-agent teams. The proposed framework addresses the main challenges associated with decentralized task execution, including high-level planning, distributed deliberation, intention sharing, conflict resolution, action parallelization, and the integration between symbolic decision-making and low-level control. The architecture is organized into a set of interconnected functional blocks, each responsible for a specific stage of the control process. The planning module generates a high-level control policy for each agent according to the shared goal, the capabilities available within the team, and the logical representation of the problem. The resulting policies are implemented through Behavior Trees, which provide a modular, reactive, and interpretable structure for representing alternative action sequences and coordinating the execution of complex tasks. The planning procedure is based on a backward-chaining expansion mechanism that distributes the task among the agents according to their available capabilities. In addition to the complete planning algorithm, which generates the control policies from an initial set of empty trees, partial replanning procedures are introduced to efficiently manage changes in the team configuration or in the set of available capabilities. These procedures make it possible to update only the affected portions of the Behavior Trees without repeating the entire planning process. During task execution, the deliberation module periodically evaluates the Behavior Tree associated with each agent and generates a proposal corresponding to the action that the agent intends to execute. A priority-assignment mechanism then associates a priority value with each proposal. Different strategies are considered, including static agent priorities, aging-based mechanisms, failure-probability-based criteria, and custom state-dependent priority functions. The proposals are exchanged through a distributed communication mechanism that allows every agent to acquire a consistent representation of the intentions of the other members of the team. Based on this shared proposal set, a priority-based filtering procedure selects the actions that can be executed. The selection process resolves conflicts caused by the use of exclusive resources, mutually disruptive action effects, or the invalidation of action preconditions. At the same time, compatible proposals are preserved, allowing multiple agents to operate in parallel and improving the overall efficiency of the team. The selected proposals are then converted into executable subtasks and transmitted to the low-level controllers. These controllers compute the physical control inputs required to perform the selected capabilities. At the beginning of each deliberation cycle, the physical state of the robots and the environment is mapped into a logical state, which is subsequently used to update the high-level policies and generate new proposals. The proposed planning and control framework is supported by a theoretical analysis of its main properties. In particular, the thesis investigates the termination of the planning procedure, the finite-time success of the generated Behavior Trees, the finite-time achievement of the shared goal from the corresponding regions of attraction, and the goal-consistency properties of the team under the considered execution assumptions. The architecture is validated through four representative multi-agent scenarios and two complementary experimental environments. A logical simulation environment is used to evaluate the complete and partial planning algorithms and to measure the ideal performance of the high-level control structure. A physics-based simulation environment reproduces the complete control pipeline, including deliberation, communication, coordination, state abstraction, low-level actuation, and robot-environment interaction. The experimental analysis considers planning time, tree size, execution success rate, task completion time, parallel execution time, and agent utilization, demonstrating the applicability of the proposed framework to heterogeneous multi-agent cooperation tasks.
Institution: University of Pisa
Language: Italian
Open Access status: green
OpenAlex ID: W7210460802
Cites: 16
Cited by:
Related to: 10