Deliberative Perception for Warehouse Automation

We describe the application of a recently developed deliberative perception framework to the task of multiobject instance recognition in warehouse environments. Traditional object recognition pipelines based exclusively on discriminative feature-matching and/or statistical learners are often sensitive to inter-object occlusions and the training data used. Deliberative approaches such as PERCH treat multi-object pose estimation as a generative global optimization over possible configurations of objects, thereby predicting and accounting for occlusions. Further, D2P—an extension of PERCH, leverages guidance from modern learning-based techniques to combine the efficiency of discriminative approaches with the robustness provided by global reasoning. We conclude with a discussion of how these approaches were used by Carnegie Mellon University’s team in the 2016 Amazon Picking Challenge, and their role in the upcoming 2017 Amazon Robotics Challenge.