• flexible structures
• learning control
• modeling planning and control for sofr robots
• motion control
• optimal control

The collective imagination of a robot is broad and subjective. For example, if we asked an adult and a teenager ‘‘What is a robot?’’ their answers would be very different. The former would probably imagine a rigid, bulky, scary, and industrial manipulator acting in a factory line; the latter would most likely picture a collaborative one with bio-inspired features operating in an unstructured scenario. Nowadays, considering robots confined to structural warehouses and factory lines is extremely disheartening and devoid of purpose. % is extremely depressing and pointless.
However, approaching unstructured, everyday life or natural scenarios opens new challenges in design, planning, and control algorithms for robotic systems.

Taking inspiration from nature, we observe that three are the main principle that governs the behavior of dynamic systems such as living beings and natural phenomena. First, most of the systems are underactuated. This means that some inputs lead to direct effects, and others have hard-to-predict consequences in the time evolution of the system. Second, living beings exploit the mechanical compliance and underactuation of their bodies to optimize energy consumption, and exploit performances. Third, living beings can modify their actions based on previous experience and knowing their strengths and weaknesses. More specifically, humans are exceptional at reshaping their skills by repeating the same ones over and over.

This thesis proposes planning and control algorithms grounded in these three natural principles. The main goal is to unlock the full potential of soft robots acting in unstructured environments. Two are the main classes of soft robots: articulated and continuum. The former takes inspiration from the vertebrate world of the living kingdom, i.e., mammals, while the latter mimics the invertebrate living beings, i.e., worms, plants, etc. Since nature does not show discontinuity in evolution, in this thesis, a third class, at the intersection between the aforementioned two, will be exploited, namely compliant underactuated robots.

Since no golden rule is present in modeling soft robots - especially in the soft continuum case - the first question is to derive suitable models for solving the planning and control problem. The solution relies on deriving a control-oriented method through the discretization procedure of the continuum bodies taking into account the elasticity and the underactuation of the robots. Not surprisingly, in such a scenario, all the deformations of the continuous body cannot be independently guided toward the desired ones. Therefore, one has to select a combination of state variables, the so-called output function, which plays a crucial role in the control problem. It is worth highlighting that underactuation is not the engineer's choice but rather something that should be exploited to unlock the maximum mechanical potential of the robot.

The second step to exploit the soft body potentialities is to plan the desired motion relying on the description of the robot. This is done by implementing a feasibility-driven multiple-shooting approach that can take into account the full soft robot's dynamics encompassing its elasticity and the underactuation.

The third step regards the control design. The controller has to preserve the elasticity of the manipulator, leverage partially the model knowledge, exploit the mechanical underactuation, and compensate for the modeling uncertainties. This paradigm translates into a learning control architecture that is mostly feedforward and that uses previous experimental data and model-based information to refine the control action until a suitable threshold of the error is reached, iteratively. Leveraging partial model information, the proposed iterative framework converges piece-wisely to the desired motion, compensates for the reality gap, and is sped up by any knowledge of the model.

A practical application of soft robots is presented, showcasing a fishing rod robot that leverages the continuum beam elasticity to precisely throw and retrieve small objects, i.e., sensors, reaching meters away target. This study represents a concrete application in which soft robots could outperform classic rigid manipulator by leveraging their elasticity while operating in a natural environment.