Talk title: Robot learning of manipulation skills by exploiting variable impedance
Abstract: Robot learning can exploit variable impedance to interact with the surrounding users and the robot's environment. In contrast to other machine learning applications, robot learning typically relies on only few demonstrations or trials. Thus, the main challenge is to find structures and priors that can be used in a wide range of tasks. To reduce the amount of required data, a first opportunity to seize is that manipulation skills acquisition in robotics is a scaffolding process rather than a standard machine learning process. Indeed, we can exploit a number of interactive learning mechanisms to acquire/generate better data on-the-spot, by relying on social mechanisms to transfer skills more efficiently, including active learning and bilateral interactions. For this purpose, variable impedance is a key component to guide the robot through the acquisition of the task by kinesthetic teaching.
To encode skills in a compact and modular manner, movement primitives can also be used in robot learning as high-level "bricks" of motion that can be re-organized in series and in parallel. To enable variable impedance, I will show that this notion can be extended to behavior primitives, forming a richer set of time-dependent and time-independent behaviors. I will propose to formalize the combination of behavior primitives as a product of experts (PoE), a machine learning technique modeling a probability distribution by combining the output of several simpler distributions. This fusion approach allows robots to counteract perturbations that have an impact on the fulfillment of the task, while ignoring other perturbations. This formulation creates bridges with research in biomechanics and motor control, including minimal intervention principles, uncontrolled manifolds or optimal feedback control.
To facilitate the acquisition of manipulation skills, task-parameterized models can also be exploited to take into account that the robot motion typically relates to objects, tools or landmarks in the robot's workspace. The approach consists of encoding a motion in multiple coordinate systems (e.g., from the perspectives of different objects), in the form of trajectory distributions. In a new situation (e.g., for new object locations), the reproduction problem corresponds to a fusion problem, where the variations in the different coordinate systems are exploited to generate a movement reference tracked with variable gains, providing the robot with a variable impedance behavior that automatically adapts to the precision required in the different phases of the task. For example, in a pick-and-place task, the robot will be stiff if the object needs to be reached/dropped in a precise way, and will remain compliant in the other parts of the task.
The last part of my presentation will discuss how variable impedance skills transfer can exploit stiffness and manipulability ellipsoids, in the form of geometric descriptors representing the skills to be transferred to the robot. As these ellipsoids lie on symmetric positive definite manifolds, the use of Riemannian geometry will be proposed as a way to learn and reproduce these descriptors in a probabilistic manner.