Energy Reduction of Robot Stations with Uncertainties

Abstract: This thesis aims to present a practical approach to reducing the energy use of industrial robot stations. The starting point of this work is different types of robot stations and production systems found in the automotive industry, such as welding stations and human-robot collaborative stations, and the aim is to find and verify methods of reducing the energy use in such systems. Practical challenges with this include limited information about the systems, such as energy models of the robots; limited access to the stations, which complicates experiment and data collection; limitations in the robot control system; and a general reluctance by companies to make drastic changes to already tested and approved production systems. Another practical constraint is to reduce energy use without slowing down production. This is especially challenging when a robot station contains stochastic variations, which is the case in many practical applications. Motivated by these challenges, this thesis presents an offline method of reducing the energy use of a production line of welding stations in an automotive factory. The robot stations contain stochastic uncertainties in the form of variations in the robot execution times, and the energy use is reduced by limiting the robot velocities. The method involves collecting data, modeling the system, formulating and solving a nonlinear and stochastic optimization problem, and applying the results to the real robot station. Tests on real stations show that, with only small modifications, the energy use can be reduced significantly, up to 24 percent. The thesis also contains an online method of controlling a collaborative human-robot bin picking station in a robust and energy-optimal way. The problem is partly a scheduling problem to determine in which orders the operations should be executed, and a timing problem to determine the velocities of the robots. A particular challenge is that some model parameters are unknown and have to be estimated online. A multi-layered control algorithm is presented that continuously updates the operation order and tunes the robot velocities as new orders arrive in the system. Simultaneously, a reinforcement learning algorithm is used to update estimates of the unknown parameters to be used in the optimization algorithms.