Data-Driven Synthesis of Perception Pipelines via Hierarchical Planning

Robotic systems in production environments have to adapt quickly to new situations and products to enable customization and short product cycles. This is especially true for the robot perception, which is sensitive to changes in environment and task. Therefore, we present an approach to quickly synthesize perception pipelines based on hierarchical planning. We calibrate the hierarchical model, such that it reflects condensed experience from historical data. We validate our approach in a simulated assembly scenario with objects from the Siemens Robot Learning Challenge, taking into account different possible sensor types and placements.