Predictive Context-Based Adaptive Compliance for Interaction Control of Robot Manipulators

In classical industrial robotics, robots are concealed within structured and well-known environments performing highly-repetitive tasks. In contrast, current robotic applications require more direct interaction with humans, cooperating with them to achieve a common task and entering home scenarios. Above all, robots are leaving the world of certainty to work in dynamically-changing and unstructured environments that might be partially or completely unknown to them. In such environments, controlling the interaction forces that appear when a robot contacts a certain environment (be the environment an object or a person) is of utmost importance. Common sense suggests the need to leave the stiff industrial robots and move towards compliant and adaptive robot manipulators that resemble the properties of their biological counterpart, the human arm. This thesis focuses on creating a higher level of intelligence for active compliance control methods applied to robot manipulators. This work thus proposes an architecture for compliance regulation named Predictive Context-Based Adaptive Compliance (PCAC) which is composed of three main components operating around a 'classical' impedance controller. Inspired by biological systems, the highest-level component is a Bayesian-based context predictor that allows the robot to pre-regulate the arm compliance based on predictions about the context the robot is placed in. The robot can use the information obtained while contacting the environment to update its context predictions and, in case it is necessary, to correct in real time for wrongly predicted contexts. Thus, the predictions are used both for anticipating actions to be taken 'before' proceeding with a task as well as for applying real-time corrective measures 'during' the execution of a in order to ensure a successful performance. Additionally, this thesis investigates a second component to identify the current environment among a set of known environments. This in turn allows the robot to select the proper compliance controller. The third component of the architecture presents the use of neuroevolutionary techniques for selecting the optimal parameters of the interaction controller once a certain environment has been identified.