Noah Wiederhold will speak on Studying Human-Human Handovers for Large Collections of Everyday Use Objects
Abstract:
Understanding how multiple people perform close interactions such as object handover is important to inform robotic systems on conducting in human-robot collaboration while being aware of human preferences for engaging in the collaborative activity. Currently, large-scale propagation of human-human handover research to generalized robot systems is hindered in three ways: the small object counts and diversity used in existing studies, the use of restrictive and low-resolution markered body motion capture technology, and the lack of participant preferences during handover. This thesis presents (a) a large human-human handover study of recordings of 32 participants handing 204 diverse objects and providing preference feedback, (b) an analysis of agreement in giver and receiver intentions on grasping during handover, (c) an analysis of giver and receiver object orientation preferences during handoff, (d) HOH: the largest to-date dataset of markerless, multi-modal human-human handover recordings of 136 diverse objects with preference information from 40 participants, (e) a comprehensive discussion of the properties of the HOH dataset, (f) a demonstration of the usefulness of HOH to the field of robotic handover research by using it to train deep neural networks that infer human-preferred parameters of handover for four tasks of relevance to the robotic manipulation pipeline, and (g) a novel hand pose optimization algorithm that is robust to the substantial occlusion often present in multi-person handover data.
Committee:
Dr. Natasha Kholgade Banerjee (advisor)
Dr. Sean Banerjee
Dr. Christopher Lynch
Dr. Abul Khondler
Dr. Maria Kyrarini
Tuesday, May 14, 2024
10:00 AM in Snell 212
Join on Zoom – 914 3673 0077
