Geometry, Physics, and Human Knowledge as Inductive Bias in Robot Learning

Abstract

Deployment of safe and fully autonomous robots is one of the main challenges that roboticists have been tackling in the last decades. One of the key problems is to endow robots with robust adaptation capabilities such that they can successfully perform in poorly modeled and highly dynamic environments. Adaptation strongly depends on the interpolation and extrapolation capabilities of the robot, and therefore a central problem is how to improve model learning such that they generalize across a large variety of tasks, environments, and unseen conditions. To this extent, a promising approach is to introduce inductive bias (i.e., prior domain knowledge) into learning models, which comes in different flavors: geometry, physics, and human knowledge, among others. Overall, these types of inductive bias may improve not only robot generalization capabilities, but also enhance data efficiency and safety in the learning processes. Additionally, we have to consider how this prior knowledge is represented. This can be in the form of language, perception, or other modalities. For example, we can ground robot control using natural language as an inductive bias.

However, when introducing inductive bias into the robot learning models, several questions arise:

1. Which types of priors should we use?

2. How can we represent prior knowledge to leverage it optimally?

3. How do we achieve a balance between inductive bias and learning from data for best generalization?

4. How can we assess tangible benefits from injecting inductive bias in robot learning?

To address these questions, we should consider that priors may be seen as human intervention into the robot learning process and, hence, may contradict the “philosophy” of automated machine learning paradigms. Yet, inductive bias may make learning models more explainable, thus paving the way to trustable robots. Finally, generalization, safety, and sample-efficiency metrics are essential to evaluate the benefits of various kinds of inductive bias in downstream robot learning applications.

In this workshop, we invite roboticists with experience in geometry, physics, and human-in-the-loop learning, to share their ideas and success stories on leveraging inductive bias in robot learning problems. Given these different perspectives, we plan to attract the attention of different researchers from a large variety of subdisciplines of robot learning.

Where to find us

• Location: OGG, Building 260, Room 073 (260.073) – OGGB4

• Virtual attendance link: tbd

Paper Submissions

Thank you a lot for your interest in submitting to our workshop. As this is the first year that CoRL is offering workshops, we have decided not to have paper/extended abstract submissions. Nevertheless, we hope that you will be able to join our workshop either online or in-person in Auckland.
If you have any further questions, please feel free to contact us.

Sponsors