2020/07/20 by Philipp Kratzer, Niteesh Balachandra Midlagajni, Kratzer, Philipp +5
Computer Science · Engineering · #FOS: Computer and information sciences #Human Motion and Animation #Human Pose and Action Recognition #Robot Manipulation and Learning #Robotics (cs.RO)
paper · pdf · doi:10.48550/arxiv.2007.10038
openalex publication_date 2020/07/20 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28
Motion prediction in unstructured environments is a difficult problem and is\nessential for safe and efficient human-robot space sharing and collaboration.\nIn this work, we focus on manipulation movements in environments such as homes,\nworkplaces or restaurants, where the overall task and environment can be\nleveraged to produce accurate motion prediction. For these cases we propose an\nalgorithmic framework that accounts explicitly for the environment geometry\nbased on a model of affordances and a model of short-term human dynamics both\ntrained on motion capture data. We propose dedicated function networks for\ngraspability and placebility affordances and we make use of a dedicated RNN for\nshort-term motion prediction. The prediction of grasp and placement probability\ndensities are used by a constraint-based trajectory optimizer to produce a\nfull-body motion prediction over the entire horizon. We show by comparing to\nground truth data that we achieve similar performance for full-body motion\npredictions as using oracle grasp and place locations.\n