• iterative learning control
• payload generalization
• soft robotics

Learning-based controllers offer a promising approach to control robots for which accurate models are unknown.
Iterative Learning Control (ILC) is particularly effective in achieving good trajectory tracking performance in the presence of unknown models or soft robots, although it faces challenges with scalability and generalization, particularly when the desired task changes.
In this view, an open problem is to generalize the control action on a learned trajectory to different "payloads".
This thesis proposes a novel method to generalize learned control actions for executing a desired trajectory with varying payloads.
The proposed method provides a feedforward control action for a new payload, making the approach suitable for both rigid and articulated soft robots.
In particular, the method demonstrates that learning the trajectory with a set of example payloads - where the number of examples equals the number of dynamic parameters describing the object plus one - allows execution of the same trajectory with any new payload, without the knowledge of the dynamic model.
The effectiveness of the proposed algorithm is validated through simulations and real-world experiments using a rigid manipulator, the Franka Emika Panda, and a hybrid system consisting of the Panda robot equipped with a variable-stiffness actuator wrist.