Dynamic Task Sharing for Flexible Human-Robot Teaming under Partial Workspace Observability

摘要

The widespread availability of lightweight robots that may safely be operated without physical barriers to separate man and machine has paved the way to robot use in small- and medium-sized enterprises (SMEs). Given these technical foundations, the goal of advancing robots from tools to human-like teammates is a research topic that lately gains considerable attention. This thesis contributes a novel approach that particularly fosters flexible use, cost efficiency and operability by the existing workforce in line with the major requirements of partial automation in SMEs: End-users without expert knowledge on robotics are enabled to share procedural task knowledge with a robot teammate by adapting paradigms from the field of skill-based task level programming. During joint execution of previously modelled tasks, the robot is considered an equal partner of human workers - human as well as robot team members are likewise granted the authority to make dynamic, just-in-time decisions regarding the distribution of work repeatedly. This requires a high level of robot capabilities and autonomy, but therefore also allows for flexible transitions between human-robot coexistence, decoupled co-working in cooperation and close interaction in collaboration. To this end, operations from the task model are classified into categories according to their individual interaction needs and agent capabilities. An exchangeable state machine for each of these interaction categories encodes the necessary course of actions for the robot when encountering respective process steps. State machine states render the system capable of (i) understanding task progress by observing operation pre- and postconditions, (ii) executing sub-tasks itself based on a robot skill framework, (iii) delegating operations to human partners or (iv) communicating to establish mutual commitment before engaging into collaboration. Decisions in favour of an operation to go about next by following transitions in the matching state machine are made in consideration of partial workspace observability, i.e. incomplete knowledge about the state of parts and task progress: The system gets along with a lean, low-cost sensor setup only consisting of a robot-mounted eye-in-hand camera and a laser range finder to track human motion. The resulting data is fused into a human-aware world model by means of a measure for trust in stored objects. This world model enables the system to share the workspace with humans efficiently. Experiments with a simulation system that emulates dynamic human behaviour when co-working on a set of benchmark tasks show that the approach can generally speed up task execution despite these limitations in sensor use. Furthermore, preliminary human subject studies with a laboratory prototype implementation suggest that the system is promising regarding intuitive operability by non-expert users. To summon up, this thesis contributes the technical foundations, proves the feasibility of and motivates further investigations on dynamic, flexible teaming under partial workspace observability.