Learning Legible Human-Robot Handovers

Object handovers, which involve the action of giving and receiving objects, are an integral task for human-robot collaboration. Current human-to-robot handover systems rely on learning from sampled human expert trajectory data [5]. These data are usually difficult to gather, as two humans must hand over many different objects while recording the human motion with a sophisticated motion capture system. Generalization can be achieved by injecting noise into the data and forcing the learning algorithm to extract the relevant signal. We propose to devise and implement a diffusion algorithm that learns from an existing handover dataset to dramatically improve performance and generalizability.

Faculty Supervisor:

Jonathan Kelly

Student:

Partner:

Ukrainian Catholic University

Discipline:

Engineering

Sector:

Advanced Manufacturing; Artificial Intelligence; Information and Communications Technology

University:

University of Toronto

Program: