Structural Priors as Inductive Biases for
Learning Robot Dynamics

Machine learning has ushered in unprecedented advancements across a myriad of domains, from speech recognition and image processing to gaming, largely due to its ability to discover complex patterns in abundant data. However, when applied to modeling and control of robots operating in the real world, these conventional machine learning techniques confront unique challenges in terms of  (1) data scarcity, (2) generalizability, and (3) safety. 

How do we overcome these challenges to build reliable robotic systems that operate in the real-world?

Encoding known structure improves generalization, reduces sample complexity, and allows for greater interpretability. This workshop aims to foster collaboration between robotics researchers, experts in learning priors, and specialists in dynamics and control. We aim to answer outstanding questions related to learning robot dynamics by tailoring the panel session and workshop discussions around the following themes:

1. Why do we need priors for learning robot dynamics, and which types of structures, symmetries, and knowledge do we have access to?

Our speakers hail from a diverse set robotics fields—ranging from manipulation and legged locomotion to soft robotics—and will discuss the challenges of learning dynamics and the critical role of priors in achieving robust performance. These experts will highlight key structural, dynamic, and physical aspects relevant to their application area.

2. Which priors are used outside the robotics community, and how can we encode this structure into our learning algorithm?

Experts from fields such as nonlinear dynamics, control, machine learning, and computational science will share tools and methods for incorporating structural and dynamic priors into learning algorithms. They'll highlight symmetries and features of dynamical systems, offering novel insights potentially unknown in robotics.