Abstract:
This thesis investigates the problem of identifying dynamics models of individual agents of a multiagent system (MAS) and exploiting these models to shape their behavior using robots extrinsic to the MAS. While task-based control of a MAS using onboard controllers of its agents is well studied, we investigate (a) how easy it is for an observer to infer both the task and inter-agent interactions of the MAS from the agents’ motions and (b) how can the observer elicit a desired behavior in the MAS by controlling robots extrinsic to the MAS. The ability to learn individual dynamics models of an aggregated system has several applications such as learning local rules in biological swarms that give rise to emergent behavior, learning tactics of an adversarial multirobot system, and learning walking preferences of pedestrians. Likewise, the ability to shape behavior using extrinsic robots can be used to defend against an adversarial team of robots, herd farm animals towards a goal using a shepherding robot and guide humans using robots as in social navigation.

The first part of this thesis focuses on the dynamics learning problem. Our approach models agents as integrators that use reactive optimization-based controllers to synthesize inputs for mediating between their task objective and safety with respect to other agents. We develop decentralized estimators that allow an observer to learn each agent’s task provided they have full knowledge about the agent’s safety constraints and perfect measurements of its positions and velocities. Necessary identifiability conditions are developed that guarantee correct inference. The message conveyed by these conditions is that a greater number of interactions amongst agents overwhelms their dynamics with safety constraints, making them compromise on task completion. This, in turn, is what results in incorrect task inference. Since inter-agent interactions are inevitable, to circumvent this hurdle, we develop a contingency estimator that learns correct bounds on the task parameters during periods of increased interactions. Next, we take recourse to the theory of inverse optimization to develop robust estimators that can infer both tasks and constraints in the face of measurement noise and model mismatch. While we have verified these estimators in simulations, we propose to experimentally evaluate their brittleness using robots in the multirobot arena in our lab where measurement noise would be unavoidable.

The second part of the thesis focuses on modifying the MAS’s autonomous behavior by inducing interactions using robots external to the MAS. Our preliminary approach uses feedback linearization to pose this behavior shaping problem. While this has been tested on modest instances involving a few agents and one robot, this approach is prone to infeasibility and intractability with an increasing number of agents. Thus, our proposed work will focus on developing a practical mechanism to scale this to a scenario comprising of multiple agents and multiple robots. We will analyze how to decouple the multirobot-multiagent behavior shaping problem into multiple instances of one robot-one agent problems that can build on our currently developed theory. Further, we propose to do experimental validation of these algorithms using teams of heterogeneous robots in our lab.

Thesis Committee Members:
Changliu Liu, Co-chair
Katia Sycara, Co-chair
Zac Manchester
Mario Santillo, Ford Motor Company

More Information