Knowledge-Based Systems For Human-Robot Collaborative Tasks in Manufacturing Environments

While factories have started using robots that are safe enough to work alongside people, they are usually not smart enough to work with people. Those robots are commonly referred to as collaborative robots. On top of being safe by design, which makes traditional fencing unnecessary and allows for saving space and money, they could enable new tasks to be performed. Ideally, teaming up allows both human and robot workers to use their strengths to complete each other, thus having robots enhance rather than supplant human capabilities. In practice, robots often work individually due to perception and reasoning limitations, i.e., collaboration is not exploited to its full potential.

This thesis presents methods that develop awareness and more generally collaboration capabilities for robots. At the core of it, a system that takes conceptual knowledge as input and can plan/execute tasks based on a set of predefined motion primitives. Two task planning techniques were explored. Commands to the robot can be expressed in natural language if they are formatted in a specific way and in case unknown terms are used a mechanism relying on a short dialog can help solve ambiguities. Collaborative actions are represented following the pattern pre-collaboration, collaboration, and post-collaboration. Knowledge can be exported via interfaces that offer an alternative to manual modifications to the knowledge base. They also give the possibility to import CAD models of parts that need to be manipulated and export properties about them. In the case of assembly tasks, parts can be matched to indicate that they need to be put together. Experiments with a Franka Emika robotic arm demonstrate how such methods can help improve the degree of collaboration with cobots.