|CASALINO ANDREA||Cycle: XXXII |
Section: Systems and Control
Tutor: BASCETTA LUCA
Advisor: ROCCO PAOLO Major Research topic
:ALLOWING A REAL COLLABORATION BETWEEN
HUMANS AND ROBOTSAbstract:
Collaborative Robotics is emerging as one of the most active line of research in Robotics. That term indicates those group of methodologies and techniques designed to let robots work side by side to humans. Indeed, human workers should execute high cognitive tasks, like e.g. assembly operations that could be too much difficult to fully automatize, while robots have to both undertake autonomous operations and assist the humans in many ways. The combination of the human flexibility and the robots efficiency can significantly improve the production process. This level of interaction requires at least the sharing of a common space. This topic has attracted the interested of many researchers in the recent years and many controlling algorithms were developed to allow a safe coexistance of humans and robots. In this
context, tracking the human motion is paramount. Then, a safe motion controller can optimize the trajectory of the robots with the aim of dodging humans. We can state that the safe interaction of humans and robots, while performing disjoint tasks, is something achieved.
For this reason, the aim of this thesis was to study more in deep the level of collaboration.
In particular, this was done by focusing on industrial contexts, were typical applications are collaborative assemblies (or co-assemblies). In such scenarios, humans and robots have to execute alternating tasks, with the aim of realizing a set of possible finite products. The robots have to adapt and synchronize with the humans, since the collaboration was conceived as human-centric. Indeed, it’s the human agent
that regulates the interaction. Then, robots have to interpret the human intentions as well as predict them in order to take the best actions for providing a reliable assistance. Such an interpretation is possible only through increased cognitive capabilities. To this purpose, sensors can be exploited to produce a big amount of data describing the workspace surrounding a robot, which are at a second stage interpreted by machine
learning techniques. This thesis has the three following main contributions:
- propose algorithms and methodologies for inferring the current action that a human operator is undertaking, from the simple ones as for instance those for reaching tools or objects, to the more abstract ones as for instance performing a screwing. Two inferring algorithms will be proposed. The first one analyzes the motion of the hands as well as the orientation of the gaze for inferring the next reaching target of an operator adopting a Gaussian Mixture model. The second algorithm takes into account the motion of the entire body and is based on Conditional Random Fields.
- predict the actions performed by human operators in the near-far future. The proposed solution is made of two parts. One models the sequence of operations, while the other one the time durations. The first kind of modelling can be done by making use of two alternative approaches, one based upon Higher Order Markov model and the other one based on the construction of a Suffix Tree.
- optimally schedule the operations assigned to robots, with the aim of assisting the human and minimizing the inactivity times. This must be done by properly taking into account the time variability of human actions. All the developed scheduling approaches consider a particular class of Timed Petri Nets, specifically derived for describing collaborative tasks. The optimal commands to be sent to robots are extracted from a reachability tree representing many alternative evolutions of the system.
Although collaborative robots are intrinsically safe, an additional minor objective of the thesis was to investigate how to optimally control their motion in collaborative cells. This was done similarly to the scheduling above exposed, by taking into account a prediction of the human motion. All the proposed methodologies were tested in realistic co-assemblies.