Multimodal State Observation and Hierarchical Failure Handling for Reactive Task and Motion Planning in Robotics

Kurzfassung

In the future, robots are expected to assist in various domains, ranging from care-giving to planetary exploration. Integrated task and motion planning enables the robot to schedule so- phisticated action sequences to solve a certain task. However, the slightest deviation from the planned sequence may result in execution failure. To become truly autonomous, robots have to acknowledge feedback and recover accordingly. In this thesis we propose a novel formalism to observe the current world state during execution and detect deviations in order to react to uncertainties to prevent failure cases. A hierarchical failure handling system manages reactions on different levels. We introduce Reaction Templates (RTs) to dene generic rule-based activation of context-specic reaction patterns. This allows to use classical deterministic approaches for fast planning and reaction mechanisms during runtime to cope with uncertainties. We integrate feedback in two ways. General assumptions that were made implicitly during action denition are monitored at runtime. In addition we offer the possibility to explicitly specify reactive behavior in the action denition. We show two tasks on the humanoid robot Rollin' Justin. In the task of pouring a drink we monitor the implicit assumption that the bottle does not move in the robots hand after picking it up and react if the deviation is too high. This enables the robot to cope with deviations of the bottle position in its hand, which would have led to spilling without our method. The second task is the disconnection of a probe to a unit. In this example we explicitly specify that the robot shall pull the plug until it detects that it has been disconnected successfully. With this feedback, the robot shortens the disconnection time by 58% on average and is able to detect if the disconnection was unsuccessful, preventing it form damaging its hand in subsequent movements.