Automated Learning of Parameters for State and Error Classifiers in Plan-Based Robotic Task Execution

Kurzfassung

As robots become more autonomous, there is a number of issues that keep them from reaching their full potential. One of these problems is the ability to detect and react to possible failures. In order for a robot to handle failures adequately, it first needs to be able to detect deviations from the expected state of the environment and determine whether they are likely to lead to failure. While a framework for detection of and reaction to these kinds of deviations already exists, the parameters for deciding whether a reaction is necessary are specific to individual actions and need to be implemented manually, which is cumbersome. In order to enable reactive behavior for a wide range of actions, the parameters for failure detection should be obtained automatically. This thesis proposes a pipeline for automatic generation of classifiers for action effect prediction in integrated task and motion planning. With this pipeline, classifiers for individual actions are trained automatically and can then be used to detect failures at runtime. A modular design for the pipeline, which can be used with both data from real-world experiments and data generated using a physics simulation, is introduced. Different classifiers are trained and compared to identify which ones are suitable for the described use case. The data used for training is generated using a physics simulation. The performance of the classifiers and the pipeline on data from two different actions is examined and reported. With this, the capability of the pipeline to automatically train classifiers for different actions using data from a physics simulation is demonstrated. With the inclusion of proposals for future work to further generalize the individual pipeline modules, this thesis lays the groundwork for an end-to-end pipeline that can aid in improving reactivity in task and motion planning.