Composing Concurrent Reactive Robot Skills: Integration of Task Specification and Control with Symbolic Planning and CAD Information

The general trend in industrial robotics sees an expansion of applications from traditional, repetitive tasks to challenging tasks requiring advanced sensing and higher intelligence. However, despite the technological advancements in robot hardware, sensors and tools, robot adoption in industry is still overshadowed by the fear of high deployment cost. This cost is partly due to the expensive programming, especially for complex tasks that involve a higher variation in the environment and objects to manipulate, and presence of disturbances. Consequently, the traditional programming methods are often no longer a viable solution. This thesis aims to address this issue by providing a set of methods to tackle specific problems in programming robotic behaviors, i.e., ``skills'', that can perform complex tasks robustly. Common in all of these methods is the formulation of skills using constraint-based task specification. This approach specifies robotic tasks in terms of constraints as opposed to traditional approach where robots are programmed in terms of desired positions or paths. It provides two main advantages that this thesis tries to exploit: (i) real-time reactivity by incorporating sensor input (ii) ability to compose different behaviors that correspond to different aspects, such as the actual tasks, safety, and environment. As a first contribution, this thesis presents a framework to specify composition of reactive and synchronized robot skills. First, this framework provides a skill model that captures deviating behavior during non-nominal conditions. The model also introduces the notion of skill progress as a measure of skill advancement towards its final state. Second, it presents composition patterns for combining skills concurrently such that certain composed behaviors are formed. The composed behaviors include skills with synchronized progresses as well as skills capable of reacting to disturbances. As a second contribution, this thesis presents a framework that integrates skills with symbolic task planning. While the previous contribution captures the reactivity at the continuous level, this contribution extends the reactivity at the discrete level. The integrated task planner does not only compute a nominal plan to achieve the given goal, but also recomputes a new plan when failures due to unpredictable disturbances occur. This new plan is generated and translated into reactive, composed skills during runtime. The result is an improved autonomy and robustness for robot applications in non-ideal environments. As a third contribution, this thesis presents a framework to specify assembly tasks in CAD domain and to semi-automatically generate the corresponding skills. This framework leverages the information embedded in CAD data in order to infer suitable skill candidates and their appropriate parameters, given a manually specified task. The CAD domain specification offers a higher-level of abstraction, therefore reducing the needs for programmers to understand lower-level skill implementation. Finally, this thesis provides experimental evaluations of the above contributions in industry relevant lab demonstrators. These demonstrators comprise assembly tasks of different workpieces where various level of uncertainties and disturbances are present. The demonstrators also show the applications of sensor-based skills in dual-arm scenarios, where tasks have to be performed concurrently.

Publication year:2022

Accessibility:Open