-- We propose a decentralized reinforcement learning solution for multi-agent shepherding of non-cohesive targets using policy-gradient methods. This model-free framework effectively solves the shepherding problem without prior dynamics knowledge. Experiments demonstrate our method's effectiveness and scalability with increased target numbers and limited sensing capabilities. The shepherding problem in robotics exemplifies the problem of harnessing complex systems for control [1], [2]. It generally involves a group of actively controlled agents, termed herders, strategically influencing a group of passive agents, referred to as targets.

We present GoRINNs: numerical analysis-informed (shallow) neural networks for the solution of inverse problems of non-linear systems of conservation laws. GoRINNs is a hybrid/blended machine learning scheme based on high-resolution Godunov schemes for the solution of the Riemann problem in hyperbolic Partial Differential Equations (PDEs). In contrast to other existing machine learning methods that learn the numerical fluxes or just parameters of conservative Finite Volume methods, relying on deep neural networks (that may lead to poor approximations due to the computational complexity involved in their training), GoRINNs learn the closures of the conservation laws per se based on "intelligently" numerical-assisted shallow neural networks. Due to their structure, in particular, GoRINNs provide explainable, conservative schemes, that solve the inverse problem for hyperbolic PDEs, on the basis of approximate Riemann solvers that satisfy the Rankine-Hugoniot condition. The performance of GoRINNs is assessed via four benchmark problems, namely the Burgers', the Shallow Water, the Lighthill-Whitham-Richards and the Payne-Whitham traffic flow models. The solution profiles of these PDEs exhibit shock waves, rarefactions and/or contact discontinuities at finite times. We demonstrate that GoRINNs provide a very high accuracy both in the smooth and discontinuous regions.

We present a data-driven control architecture for modifying the kinematics of robots and artificial avatars to encode specific information such as the presence or not of an emotion in the movements of an avatar or robot driven by a human operator. We validate our approach on an experimental dataset obtained during the reach-to-grasp phase of a pick-and-place task.

A key problem toward the use of microorganisms as bio-factories is reaching and maintaining cellular communities at a desired density and composition so that they can efficiently convert their biomass into useful compounds. Promising technological platforms for the real time, scalable control of cellular density are bioreactors. In this work, we developed a learning-based strategy to expand the toolbox of available control algorithms capable of regulating the density of a \textit{single} bacterial population in bioreactors. Specifically, we used a sim-to-real paradigm, where a simple mathematical model, calibrated using a few data, was adopted to generate synthetic data for the training of the controller. The resulting policy was then exhaustively tested in vivo using a low-cost bioreactor known as Chi.Bio, assessing performance and robustness. In addition, we compared the performance with more traditional controllers (namely, a PI and an MPC), confirming that the learning-based controller exhibits similar performance in vivo. Our work showcases the viability of learning-based strategies for the control of cellular density in bioreactors, making a step forward toward their use for the control of the composition of microbial consortia.

In addressing control problems such as regulation and tracking through reinforcement learning, it is often required to guarantee that the acquired policy meets essential performance and stability criteria such as a desired settling time and steady-state error prior to deployment. Motivated by this necessity, we present a set of results and a systematic reward shaping procedure that (i) ensures the optimal policy generates trajectories that align with specified control requirements and (ii) allows to assess whether any given policy satisfies them. We validate our approach through comprehensive numerical experiments conducted in two representative environments from OpenAI Gym: the Inverted Pendulum swing-up problem and the Lunar Lander. Utilizing both tabular and deep reinforcement learning methods, our experiments consistently affirm the efficacy of our proposed framework, highlighting its effectiveness in ensuring policy adherence to the prescribed control requirements.

A significant challenge in control theory and technology is to devise agile and less resource-intensive experiments for evaluating the performance and feasibility of control algorithms for the collective coordination of large-scale complex systems. Many new methodologies are based on macroscopic representations of the emerging system behavior, and can be easily validated only through numerical simulations, because of the inherent hurdle of developing full scale experimental platforms. In this paper, we introduce a novel hybrid mixed reality set-up for testing swarm robotics techniques, focusing on the collective motion of robotic swarms. This hybrid apparatus combines both real differential drive robots and virtual agents to create a heterogeneous swarm of tunable size. We validate the methodology by extending to higher dimensions, and investigating experimentally, continuification-based control methods for swarms. Our study demonstrates the versatility and effectiveness of the platform for conducting large-scale swarm robotics experiments. Also, it contributes new theoretical insights into control algorithms exploiting continuification approaches.

One of the major challenges in Deep Reinforcement Learning for control is the need for extensive training to learn the policy. Motivated by this, we present the design of the Control-Tutored Deep Q-Networks (CT-DQN) algorithm, a Deep Reinforcement Learning algorithm that leverages a control tutor, i.e., an exogenous control law, to reduce learning time. The tutor can be designed using an approximate model of the system, without any assumption about the knowledge of the system's dynamics. There is no expectation that it will be able to achieve the control objective if used stand-alone. During learning, the tutor occasionally suggests an action, thus partially guiding exploration. We validate our approach on three scenarios from OpenAI Gym: the inverted pendulum, lunar lander, and car racing. We demonstrate that CT-DQN is able to achieve better or equivalent data efficiency with respect to the classic function approximation solutions.

Geometric pattern formation is crucial in many tasks involving large-scale multi-agent systems. Examples include mobile agents performing surveillance, swarm of drones or robots, or smart transportation systems. Currently, most control strategies proposed to achieve pattern formation in network systems either show good performance but require expensive sensors and communication devices, or have lesser sensor requirements but behave more poorly. Also, they often require certain prescribed structural interconnections between the agents (e.g., regular lattices, all-to-all networks etc). In this paper, we provide a distributed displacement-based control law that allows large group of agents to achieve triangular and square lattices, with low sensor requirements and without needing communication between the agents. Also, a simple, yet powerful, adaptation law is proposed to automatically tune the control gains in order to reduce the design effort, while improving robustness and flexibility. We show the validity and robustness of our approach via numerical simulations and experiments, comparing it with other approaches from the existing literature.

EXTENDED ABSTRACT Model-free reinforcement learning (or simply reinforcement learning, RL, in what follows) is increasingly used in applications to solve a wide variety of control problems (Kober et al., 2013; Garcıa and Fern andez, 2015; Cheng et al., 2019). The lack of requiring a formal model of the plant renders it appealing for a heuristic, low-cost control design approach that can be easily implemented and adapted to different situations. As a tradeoff, learning processes often require a long training phase where the controller agent learns by trial-and-error how the plant responds to different control actions, and what actions to take to steer its behavior in a desired manner. This problem is particularly relevant when using tabular methods, such as Q-learning, in those situations where reinforcement learning is applied to control dynamical systems defined in continuous spaces (Lillicrap et al., 2019). It is therefore desirable to enhance the learning process by encoding some qualitative knowledge of the system dynamics via appropriate models.