Data-efficient Robot Learning from Human Demonstrations via Interaction Graph

The proposed research project aims to solve a critical challenge in robotics, the need of large amounts of training data, which is often costly and time-consuming to collect. Interaction graphs have shown impressive generalization capability by retargeting human motions to robots from just a single demonstration. However, the behaviors exhibited from existing frameworks leveraging interaction graphs are mostly unimodal. Leveraging Diffusion Policy’s expressiveness in modelling multimodal behaviors, we propose a unified framework that combines interaction graph and diffusion policy to learn diverse manipulation strategies from just a handful of human demonstrations. This collaboration benefits both institutions by advancing the research of both labs. This project expands Prof. Allen’s prior work on interaction graphs as a robust abstraction tool to model human-object and robot-object interactions, while simultaneously supporting Prof. Coros’ focus on efficient robotic skill learning by demonstrating how structured representations such as interaction graphs can dramatically improve the sample efficiency of learning real-world tasks.

Faculty Supervisor:

Kelsey Allen

Student:

Partner:

ETH Zurich

Discipline:

Computer science

Sector:

Education

University:

The University of British Columbia

Program: