Safety and adaptation in physical interaction control for robotic applications

Abstract

Although robotic platforms have been used for a wide range of purposes, recent advances in robot autonomy raise the opportunity of bringing them closer to humans. This implies that, under any circumstance, robots have to remain safe --i.e. not harm humans, the environment or themselves-- while embedding the required adaptation means to deal with the unstructured nature of anthropic domains. Hence, initial solutions delivered reactive behaviours in order to avoid collisions, which is not suitable for those tasks where the robot is required to initiate, maintain or regulate contact over physical interaction with a human, an environment or both. At control level, which represents the earliest reaction mechanism, this calls for the development of solutions that address the new sources of hazard in physical interactions that might jeopardise robot safety while providing the required level of adaptation. This is the starting point of this Thesis, focused on the use of advances in control theory to benefit safety and adaptation in solutions for physical interaction tasks. For this purpose, recent state-of-art control solutions are analysed through a novel architecture conceived for physical interaction tasks. From the perspective of safety and adaptation concepts, it shows an absence of a systematic framework to set guarantees on operation beyond stability. Moreover, existing approaches come at the cost of restrictions on adaptation strategies, which might deteriorate robot performance, or increase its complexity, which hinders their deployment. Hence, this Thesis proposes the use of a model-based approach using Linear Parameter Varying (LPV) paradigm to formulate solutions and systems in combination with the formulation of conditions in terms of Linear Matrix Inequalities (LMI). The LPV representation provides linear-like descriptions through a set of varying parameters. In such way, a depiction of its complete range of operation can be defined by only considering the limits of these parameters. Thus, stability and other conditions, e.g. limits on control variables, can be imposed through their representation as LMIs. For both design and analysis, multiple LMIs can be simultaneously introduced as constraints into convex optimization problems, which are computationally tractable. Therefore, this Thesis presents methods addressing different problems in robotic manipulators, all of them using the LPV-LMI framework. First one consists of a compliant joint-level state-feedback controller considering robot dynamic model and introducing adaptation through a gain-shifting mechanism, which is embedded using the LPV formulation. It is presented in combination with a Robust Unknown Input Observer to estimate interaction force under a lack of state measurements, which is designed using a novel design introducing optimal conditions to improve its noise reduction capabilities, thanks to the modularity of LMIs. The second method is an automated off-line algorithm to tune Variable Impedance Controllers at task-level such that conditions regarding safety and performance are guaranteed during its operation. It is applied for both ad-hoc defined modulation laws and extracted ones from user demonstrations on the task. All methods are applied to real robotic platforms performing different physical interaction tasks. Obtained results show that the LPV-LMI framework represents a promising systematic approach to ensure a safe behaviour on different control solutions with minimal limitations to adaptation strategies in physical interaction tasks.