Robotic assembly is mainly used inside factories where both the environment and the task for each robot stays constant and the batch sizes are large, with car factories presenting a prime example. However, in manufacturing Small and Medium-sized Enterprises (SMEs) or construction yards the level of automation is very low, mainly due to the changing environment causing two major problems for robots: firstly, the programming of robots is often difficult and thus it can take too long to make the same robot perform multiple tasks interchangeably. Secondly, the use of robots with traditional control methods requires an accurate model of the environment, which can be either costly to acquire and prone to accidental changes in the real environments SMEs or simply infeasible (construction). To enable the use of robots in new environments, robots must be easy to teach and able to adapt to small changes in the environment. In this thesis we propose to use Learning from Demonstration (LfD) with compliant motions to overcome the aforementioned problems. In LfD the user can show the robot how to perform a required task, using either teleoperation or kinesthetic teaching where the teacher physically holds a gravity-compensated robot and leads it through the desired task. We developed methods to ease the use of compliance on three different levels in programming a robot: on the control level, on the primitive level and on the motion sequencing level. On the control level, we propose using either direct force control for cases where either the robot or the object is free-floating, and impedance control for cases where both the manipulator and object are ground based. On the primitive level we present a new impedance control-- based motion primitive which can be used to learn and encode motions that use the environment to mitigate pose uncertainties. Humans naturally have the skill to exploit contact forces in insertion tasks, and we want to convey this skill from human to robot in an efficient way. On the motion sequencing level we first show how a complex human demonstration can be segmented into phases, each of which can be modeled with the primitive. Then we present how the primitives can be sequenced online to successfully reproduce the task. Additionally, we show that the presented motion primitive can also be applied effectively for bimanual assembly tasks. Finally, we present how to learn from human teachers search motions similarly as a human inserting a plug into a socket in darkness, which can be used as efficient exception strategies in assembly. To conclude, this thesis present a framework that can accelerate the degree of automation in tasks where currently the use of robots is infeasible.
|Translated title of the contribution||Robottien joustavien kokoonpanotaitojen oppiminen ihmisen esimerkistä|
|Publication status||Published - 2019|
|MoE publication type||G5 Doctoral dissertation (article)|
- learning from demonstration
- compliant motions
- impedance control