If you are interested in one of the below projects, Semester Projects, Master Projects at EPFL, or Master Projects in Industry please contact the first person of reference indicated in each description either by telephone, or by email, or by visiting us directly to the LASA offices.
Robot Latent Space Acquisition for Model Learning and Robust Adaptation
Accurate robot control, specially in torque, relies on either precise knowledge of the robot dynamics or the ability to rapidly adapt to disturbances. In practice, uncertainties such as the links' mechanical properties, joint friction, or measurements' noise make acquiring and estimating such a model challenging. To address this problem, there are several approaches both offline and online: like supervised machine learning techniques (SVR, LWPR, Neural Networks ...), or online parametric model estimation . Even though these strategies are effective, they suffer from the large dimensionality of the input space (number of joints x 3), the sensitivity to measurement noise, and persistent excitation. In such cases, offline accurate model learning becomes demanding to acquire and online adaptive prediction will require rigorous gain tuning. To deal with these problems, one solution is to perform model learning or adaptation in a lower-dimensional space, the latent space, where the dimension of the input and the noise will be reduced.Thereby, one could ease the learning process and enhance accuracy control. Finding such a latent space attenuates the well-known excitation issue, and allows us to utilize high-accuracy methods which were suffering from curse of dimensionality, like fast and robust adaptation to unknown nonlinearities and perturbations. This project aims to explore this direction by testing various dimensionality reduction techniques to improve and ease the offline/online learning of the robot dynamics.