Progressive learning and its application to robot impedance learning

摘要

An approach to learning control using an excitation scheduling technique is developed and applied to an impedance learning problem for fast robotic assembly. Traditional adaptive and learning controls incur instability depending on the reference inputs provided to the system. This technique avoids instability by progressively increasing the level of system excitation. Called progressive learning, it uses scheduled excitation inputs that allow the system to learn quasistatic parameters associated with slow input commands first, followed by the learning of dynamic parameters excited by fast input commands. As learning progresses, the system is exposed to a broader range of input excitation, which nonetheless does not incur instability and unwanted erratic responses. In robotic assembly, learning starts with a slow, quasistatic motion and goes to a fast, dynamic motion. During this process, the stiffness terms involved in the impedance controller are learned first, then the damping terms and finally by the inertial terms. The impedance learning problem is formulated as a model-based, gradient following reinforcement learning. The method allows the suppression of excessive parameter changes and thereby stabilizes learning. By gradually increasing the motion speed command, the internal model as well as the control parameters can be learned effectively within a focused, local area in the large parameter space, which is then gradually expanded as speed increases. Several strategies for motion speed scheduling are also addressed.