DAY 2: RBCCPS Session

Insect nervous systems have evolved to make useful reductions of sensory-rich high-dimensional data that are suitably parsed to generate a handful of motor commands for regulating flight behavior. Understanding these processing mechanisms could then be the key to developing feedback control architectures that enable rapid decision-making in uncertain and dynamic environments given limited computational resources and noisy sensory data. To this end, multiple bio-inspired strategies for sensorimotor control are presented that leverage tools from traditional control theory to achieve safe collision-free navigation in GPS-denied unstructured environments. Numerous validation studies of aerial, ground and underwater systems are presented that attest to the efficacy of the proposed schemes. These results and their implications for the realization of autonomous systems suitable for real-world applications will be discussed.

Systems with large populations of agents occur both naturally and in engineering applications. Examples include fleet redistribution of ride sharing services, evolution of transportation mode choices of a population, opinion dynamics, human and insect swarm migrations and robotic swarms. In many of these applications, there are restrictions on the revisions of the agents’ choices depending on the population state. Such restrictions can be modeled by a network or graph of choices. In this talk, I will present some of our recent work on the dynamics of a population of optimum seeking agents on a network of choices. Specifically, we will see population dynamics under varying levels of coordination – from selfish agents to fully coordinating agents. I will summarize our results on existence and uniqueness of solutions, asymptotic convergence to the set of nash equilibria, sufficient conditions on uniqueness of nash equilibrium and an algorithm for efficiently obtaining meaningful bounds on the steady state social utility.

Security concerns arise in all sectors of power systems, from generation to distribution, control, and consumption. In this talk, I will present our recent work on dynamical load-altering attacks in power networks, where an attacker aims to destabilize the network by modifying a portion of the loads. Dynamic load-altering attacks can be implemented by tampering with demand response and demand side management services. These attacks act as sparse perturbations to the network matrices that alter key dynamical properties, thus constituting a form of distributed or sparse controller for network systems. We cast the problem of designing minimally-invasive load-altering attacks as a sparse stability radius optimization problem, and present a numerical algorithm to efficiently compute the optimal attacks. This analysis provides a vulnerability map that identifies secure and vulnerable loads in the power network.

The goal of this talk is to understand the principles of Model Predictive Control (MPC) and Reinforcement Learning (RL) in the context of legged locomotion, and investigate their strengths and weaknesses. MPC based controllers are fast and robust to a broad class of model based uncertainties. However, they do not scale well in realization of more complex locomotion behaviors. Similarly, RL based policies are able to automatically achieve complex locomotion on a variety of terrains. However, they are slow and very sensitive to the trained model. Therefore, the goal of this talk is to develop a unified framework that includes the best from both of these methodologies. We will show how MPC can help accelerate RL, and vice versa, thereby developing robust policies for a variety of benchmark locomotion tasks.