Learning to Learn for Robotics

Recent years have seen a lot of interest in the development and use of learning-to-learn for a wide range of applications. Yet, existing solutions rarely take into account the structured nature of a learner’s body which constrains their applicability to learning in physical environments, an aspect which is core in robotics.

In this workshop, we’d like to discuss how humans use their own embodied structure to learn effectively and efficiently, and what we can do to enable robotic learn-to-learn in a similar fashion. Our goal is to bring together researchers from a variety of backgrounds with the hope to discuss and reason about what learning-to-learn and embodiment means from a cognitive neuroscience perspective, and how this knowledge might translate into improving robot learning.

Concretely we are interested in the following questions:

1. What can we learn from humans utilising their bodies in the process of learning-to-learn;

2. How to make use of the embodied structure of a robot to improve or self-supervise the process of meta-learning; do we need new algorithms for embodied meta-learning?

3. Do people learn-to-learn different concepts separately or do we learn continuously throughout our lifespans; how do we make sense of the surrounding environment in this process?

4. How can we ensure learning-to-learn is safe for an embodied robot and its environment?

5. We also solicit submissions on negative results on meta learning that help us understand the current limitations and boundaries of learning to learn

Submission Form: https://forms.gle/Y6yMiSY3tc3wmF969

Submission Deadline: May 15th 2021 Deadline extended to May 28th 2021

Notification: May 31st 2021

Camera Ready Submission: tba

Workshop: June 4 2021

for formatting instructions please refer to https://www.ieee.org/conferences/publishing/templates.html

papers should be maximum 3 pages long excluding references. We accept work that has been submitted but not yet published at another conference.

Our plan is to collect novel contributions that showcase how lifelong and meta-learning strategies are applied on physical robotic tasks. We welcome submissions on but not restricted to the following topics:

Topical Areas

• Data-efficiency via transfer/multitask/meta learning for real world tasks

• Embodied learning-to-learn

• Robot learning through embodied self-supervision

• Online / fast adaptation to changing dynamics models and other sources of covariate shift

• Domain/Policy adaptation between (for example) different robotic platforms, collaborative settings, varying environments and tasks

• Standardised frameworks, metrics for physical evaluation etc. for learning to learn algorithms

• Causality in the context of Learning to learn for Robotics.

• Learning to learn with physical safety for the robot and its environment.

• Social learning-to-learn by imitating other agents’ learning strategies.

We seek to collect interesting ideas and applications addressing problems across these topics which can provide an informative overview of the existing challenges in using meta learning approaches to robotics applications and can set the grounds for future development.