Fast Learning-Based Motion Planning and Task-Oriented Calibration for a Humanoid Robot

Kurzfassung

Dextrous humanoid robots hold immense potential from manufacturing to healthcare. The capability to move and act autonomously in challenging environments is critical for their successful application. Besides understanding the environment, the robot needs an accurate model of itself to navigate safely and perform tasks accurately. The thesis tackles this interplay of fast and accurate motion planning for humanoid robots. The first part focuses on accurately modeling the robot's kinematics, including elasticities and mass distribution, through efficient and self-contained calibration. We introduce a general approach that works for diverse robots and calibration setups. Our theory strictly distinguishes between the actual measurement setup used to collect the data and the robot's intended task, which should be improved through calibration. A typical example is using a camera to measure pixels to improve an end-effector's cartesian accuracy. In the trade-off between a minimal and self-contained measurement setup and the full description of all aspects of the task, this work combines a probabilistic view of the data collection with minimizing the intended task error. The methods are demonstrated for the elastic humanoid Agile Justin using its internal RGB camera and for the DLR Hand-II with its tree-like structure using contact measurement, highlighting the viability on complex hardware and the broad applicability of the approach. Avoiding self-collision and handling obstacles in diverse environments are critical issues for motion planning. For a complex robot in an unstructured world, motion planning has many local minima of various quality and can have drastically different solutions if the input (i.e., world, start, target) changes only slightly. We use an optimization-based technique for its speed and easy extensibility. However, the cost landscape is challenging for gradient-based techniques, as they operate only locally and, thus, depend entirely on the initialization to avoid local minima. The idea is to mitigate this dependence by using neural networks to predict educated initial guesses, which lie in the area of attraction of the global minimum. We discuss two variants for training such networks. A supervised approach which uses an exhaustive dataset of successful motions in hallenging environments, and an unsupervised approach which uses the objective function directly to update the network weights. Crucial for training and the ability to generalize to new environments was to encode the worlds with Basis Point Sets. By extending the work to inverse kinematics, additional insight could be achieved into the network structure in the context of the mode switches, which are fundamental for motion planning. Overall, the work improved the accuracy and planning time of the humanoid Agile Justin with its 19 degrees of freedom. The maximal error at the end-effector was reduced from 6 cm to 0.8 cm, and the motion planning time in self-acquired high-resolution voxel models from up to 10 s to realtime-capable 0.2 s. Both improvements are significant steps towards more dexterity and autonomy and are meanwhile indispensable for the daily work on the research robot in the lab and at multiple fairs and conferences.