We are exploring ways to design control protocols for large-scale network systems that not only allow the network to achieve some desired behavior but also to reject polynomial disturbances and to guarantee the non-amplification of other disturbances across the system. We call this property network scalability and the approach we are pursuing to guarantee scalability is through the design of multiplex network layers delivering suitable integral actions.

One application involves the design of control protocols guaranteeing that a network of robots is able to: (i) fulfill a desired configuration; (ii) reject polynomial disturbances and prohibit the amplification of non-polynomial disturbances. We have implemented and validated our algorithms in a real robotic test-bed via the Robotariun by GeorgiaTech. In the experiments, two robots were perturbed. The experiments in the video show the effectiveness of our theory!

If you want to know more about our results on scalability of network systems, see e.g., here and here!

Author of the code and simulations: Shihao Xie. The code can be found on our Github page. See the gitub for more videos, details and a poster presented at NecSys2022 illustrating the results.

We have been exploring the use our latest findings in data-driven control and crowdsourcing to design autonomous agents that are able to execute complex tasks by using data from third parties.

A first application involves the design of data-driven control algorithms for autonomous agents system for connected and autonomous vehicles that...manages car parks in our UNISA Campus. Our campus is effectively a mini-city...and its parkings can become full during rush hours! With our system, cars are now able to find an optimal parking lot that takes into account the data from the surrounding environment...simulation results are encouraging, especially when the service is implemented on a network of vehicles!

A second application involves the design of autonomous agents (drones in the example) that are able to autonomously explore new, uncertain, environments by patching together data from third parties. Specifically, we are designing agents/robots able to perform complex movements from simple motion primitives....for which only a few data are available.

If you want to know more about the theory that makes this possible check out our papers here and here.

Authors of the code and simulations: Claudio Ciavola (smart cities) and Simone D'Alessio (robot exploration).

We are proud that with our recent work "A network model of Italy shows that intermittent regional strategies can alleviate the COVID-19 epidemic" we also make open tools that can be useful to evaluate both scenarios for the spreading of the Covid-19 epidemics and mitigation strategies. Our tools, with the data we used, can be found here. On the left, you see some coverage of our work from the Italian news (the article is in Italian). The work stems from a collaboration with the group of Professor di Bernardo (University of Naples Federico II), Dr. de Lellis and Lo Iudice (University of Naples Federico II), Dr. Liuzza (Enea) and Dr. della Rossa (Polytechnic of Milan).

Model-free reinforcement learning (RL) has become a popular paradigm to design autonomous agents. While its key idea of learning a policy by interacting with the environment via a trial and error mechanism is fascinating, it has two key practical disadvantages: its sample inefficiency and its lack of safety guarantees while learning. With this project we proposed a new algorithm, the MPC-augmented RL, which combines in a novel way model-based control with RL. Specifically, we explore how the availability of partial models can be leveraged to drive, and indeed accelerate, the learning phase of RL. The results are exciting: we can quickly learn a policy, while guaranteeing safety and with low sample complexity! Check out the details here or get an idea of our approach with the video on the left (it appears that cookies must be enabled to play the video)!