Machine Intelligence

One of our core research themes is the development of novel machine learning paradigms that enable complex dynamical systems with the ability to learn its self, how to interact with the world and to autonomously acquire and generalize knowledge. For this, the integration of nonlinear control theory and machine learning with intelligent data representations opens up entirely new perspectives and avenues in generating machine intelligence. Although this discipline is still in its early days there is a surprisingly high number of robotic problems that can be processed with existing results.

Contributions

Overall, our group develops machine learning for embodied systems approaches in a variety of areas. A central field has been learning tactile manipulation [160]. We introduced a skill formalism [161] that reduces the skill solution parameter space via a Parameter Space Partitioning (PSP) algorithm to a reasonable area that allows to express the solution with significantly less parameters [148]. This makes the learning of the reduced number of parameters more manageable via data-driven methods and allows to reliably learn and solve tasks such as peg-in-hole and key-in-lock with only a few trials. We have also exploited the capabilities of tactile robots to learn to detect and classify contacts [162, 163] via neural networks (NN) and other machine learning methods. The results demonstrated that NNs can be trained to detect collisions and discriminate between the collided materials, even with different human body parts [164]. Another of our research areas focuses grasp planning and learning where, in recent work [165], we introduced a segmented-mesh-based approach using objects and human knowledge as center pieces for grasping posture prediction. Our results generated feasible posture estimations in objects which are significantly different than the ones used as training samples. Furthermore, in the area of learning optimal controllers from optimal control [59] we developed supervised learning methods that use numerical solutions of optimal control problems to generate training data for learning a function approximate to an optimal controller. This Learning from Optimal Control (LfOC) approach has been successfully demonstrated for real-time quadrotor aggressive maneuvers[166] and throwing/reaching tasks for robotic manipulators [98].

Additionally, we work on the area of physics-informed machine learning, in particular we work on the translation of conventional system identification methods into a sequence of learning modules that enable robots to acquire their body model [21]. Our results have, for instance, enabled the online learning of physically feasible parameters for robot manipulators [169] or learning the entire morphology from proprioceptive sensing only [170]. Finally, we research on evolutionary robot synthesis involves the use of machine learning to enable the training of design models, which automatically generate robot parts and sub-systems to fulfill the requirements of new tasks and environmental conditions and execute a defined set of tasks time and cost efficient compared to traditional design adaptation approaches. We integrate model-based compliant control approaches with human neuromusculoskeletal models and learning algorithms [68]. Technology transfer is a core focus of our group; therefore, we work with industry and clinical partners to test our systems in the field as early as possible.