Semantic State Validation for Tele-Robotic Manipulation

Kurzfassung

Humans employ robots as tools. The purpose can range from operating in remote areas over augmenting human capabilities to freeing humans of recurrent tasks by means of autonomy. While research in robot perception is mostly focused on perceiving the geometrical state of the world, planning on the geometrical level is still infeasible due to its continuous nature. Therefore, a discretization of the geometric state of the world into a symbolic state is executed and planning is based on the symbolic state. In paradigms as the action templates [1], robots know how certain actions are supposed to change the symbolic state of the world and are able to update it accordingly. However, robots are not yet able to execute every task with human-level success rate. Additionally, in some cases, as supporting humans with limitations, human users want to control the robot in a teleoperated fashion. In order to bring together the advantages of teleoperation and autonomy, traded control is employed. In this control scheme the robot can either be teleoperated or commanded via high-level commands. Whenever the robot is teleoperated, it is commanded on a joint-level and does not know about the impact it has on the geometric and symbolic state of the world. Thus, as soon as it gets handed back control, the robot does not possess an accurate representation of the world state as needed for further autonomous planning. A solution to the given problem statement is presented by creating multiple modules to run on the robotic platform Rollin’ Justin, a humanoid robot designed and built at the German Aerospace Center. The most important modules are a physical simulation that predicts effects of actions on the environment and an inference module that infers the symbolic state of the world from the outcome of the simulation. In order not to hard-code the inference information into the inference module, python code snippets are created for holding this information. They are stored alongside other object information in the object database, making the system more flexible and adaptive to different scenarios and increasing maintainability. Furthermore an Evolutionary Strategy is implemented which allows to optimize simulation parameters with respect to the deviation between the simulated and measured geometrical state of the world, the so called reality gap. Finally it is shown how this approach can be used in order to refine and correct object knowledge during the interaction with objects.