Example of a randomly generated map with 12 obstacles, 5 cops, and 1 robber. The cops form the team of robots intercepting the target, robber.

This paper aimed to evaluate the efficacy of centralized and distributed Model Predictive Control (MPC) algorithms in a 2D adversarial control problem with planar quadrotors. The control problem is a team-based 2D simulation where a team of robots are attempting to intercept a target, and an individual robot acting as the target whose goal is to evade the team of robots for as long as possible. In the centralized form, the team of robots has its own centralized MPC algorithm while the target robot has its own MPC algorithm. For the distributed form, each robot has its own MPC algorithm, but the objective function is coupled for the team of robots and the constraints are decoupled. Furthermore, in both forms all the agents in the environment have perfect information about the location of the rest of the agents, but not their planned trajectories. These MPC algorithms were evaluated using: interception time, computation time, and capture success rate in different scenarios and environments. The goal was to compare the computational load of distributed versus centralized MPC over a variety of parameters as well as to improve the behavior of both the target and the captors. The results of experiments over a range of environments and values for the number of agents and the length of the time horizon showed that distributed MPC has a significantly lower computational load, especially when scaled up to a larger team of robots. Both the centralized and distributed algorithms performed similarly in simulated game results.

Video showing example experiments is located below.

Full Project Report Found HERE.

Note: Project conducted with Akhil Devarakonda and Rahul Zahroof.