I received the Ph.D. degree in Biomedical, Electrical, and System Engineering from the University of Bologna in 2021 and the M.Sc. degree with honors in Control Systems Engineering from the University of Padova in 2016.
From 2021 to 2023, I worked as a postdoc at the University of Bologna. Currently, I am a postdoc at the Robotics and Mechatronics (RaM) group at the University of Twente.
Follow the link to see my Curriculum Vitae
Organisaties
My research currently focuses on developing optimal control techniques for robotic systems that physically interact with unstructured environments. To achieve this, I am exploring the integration of data-driven methods, such as machine learning, with prior physical knowledge, including energy flow information and geometric structures. In parallel, I work on advancing control frameworks, enabling their effective integration with optimization approaches. Additionally, I am exploring safe physical interaction between robots and unpredictable environments, including potential interactions with humans, by equipping the system with energy-awareness to enhance its responsiveness and safety.
Recent advancements in machine learning have gained significant interest, positioning these methods as promising tools for designing controllers capable of handling complex tasks in unknown environments [1,2]. However, these approaches struggle to guarantee stability and safety, critical in human-robot interaction, and they require large amounts of labeled data, which is usually prohibitive for robotic applications. To overcome these challenges, it is crucial to encode some form of structure into the control scheme that provides prior information [3].
Energy is a fundamental concept in physics, offering intuitive insights into system dynamics. Energy-based methods leverage the structural properties of physical systems, shaping both static and transient behaviors. These methods provide a framework for understanding complex interactions by decomposing system energy into its subcomponents and interpreting the interconnections between subsystems. This is particularly important for robots interacting with unstructured environments. Port-Hamiltonian theory and passivity-based control offer a framework for designing systems that respect energy conservation and dissipation laws, providing valuable insights into energy flows that can be used to design control strategies for robots, ensuring safe and stable interactions with the environment.
By combining optimization approaches with energy-based schemes, I aim to design control strategies that balance performances and flexibility with the reliability and safety necessary for real-world tasks.
Main Challenges
- Trade-Off Between Structure and Flexibility
One challenge is finding the right balance between incorporating sufficient structure in the control scheme to guarantee stability and safety, while also leaving enough freedom in the optimization search space to enable effective controller optimization across a range of tasks and scenarios. In [5], we introduce passivity guarantees to a general-purpose control scheme based on reinforcement learning by integrating a virtual energy tank that limits energy consumption during the exploration phase. This approach enables the agent to learn a passive policy that effectively addresses the desired robotic task. In [6], we present a framework where the search space for the optimal control input is parameterized to yield a structurally passive, specifically lossless, nonlinear feedback controller. This control methodology is designed to work in conjunction with other stabilizing controllers, ensuring the stability of the closed-loop system while accommodating a broad range of optimization-based techniques, such as model-predictive control. - Integration of Optimization Approaches
Another challenge is simplifying the integration of optimization techniques, making the overall system more manageable and computationally efficient. In [4], we leverage a discrete-time formulation of port-Hamiltonian systems to facilitate the integration of control performance optimization via gradient descent on an artificial neural network, which serves as a function approximation tool. In [5], we combine deep reinforcement learning with a discrete-time implementation of virtual energy tanks to address the sampled passivity-based control problem and constructively learn passive policies for robotic applications - Understanding Physical Interactions
A key challenge lies in obtaining the necessary insights from the physical interaction between robots and their environments. In [7], using a port-Hamiltonian formalism, we provided constructive tools to study both the qualitative and quantitative effects of safety-critical control schemes implemented with control barrier functions on the energy balance of controlled physical systems. This analysis led to the development of novel energy-aware schemes, such as selective damping injection mechanisms and active control strategies, which inject energy into the controlled system to achieve the desired closed-loop behaviors. The goal for the future is to enable the system to develop a form of self-awareness. This awareness will allow the system to identify and mitigate potential risk hazards during interaction, ensuring that safety is prioritized, even in the face of unforeseen dynamics.
References
[1] Zanella, R., De Gregorio, D., Pirozzi, S. and Palli, G., 2019, April. Dlo-in-hole for assembly tasks with tactile feedback and lstm networks. In 2019 6th International Conference on Control, Decision and Information Technologies (CoDIT) (pp. 285-290). IEEE.
[2] Zanella, R. and Palli, G., 2021. Robot learning-based pipeline for autonomous reshaping of a deformable linear object in cluttered backgrounds. IEEE Access, 9, pp.138296-138306.
[3] Caporali, A., Kicki, P., Galassi, K., Zanella, R., Walas, K. and Palli, G., 2024. Deformable linear objects manipulation with online model parameters estimation. IEEE Robotics and Automation Letters.
[4] Zanella, R., Macchelli, A. and Stramigioli, S., 2024. Learning the Optimal Energy-based Control Strategy for Port-Hamiltonian Systems. IFAC-PapersOnLine, 58(6), pp.208-213.
[5] Zanella, R., Palli, G., Stramigioli, S. and Califano, F., 2024. Learning passive policies with virtual energy tanks in robotics. IET Control Theory & Applications, 18(5), pp.541-550.
[6] Zanella, R., Califano, F., Franchi, A. and Stramigioli, S., 2024. Lossless optimal transient control for rigid bodies in 3D space. arXiv preprint arXiv:2410.15984.
[7] Califano, F., Zanella, R., Macchelli, A. and Stramigioli, S., 2024. The effect of control barrier functions on energy transfers in controlled physical systems. arXiv preprint arXiv:2406.13420.
Scan de QR-code of