ICRA 2025 Workshop on Structured Learning for Efficient, Reliable, and Transparent Robots

It is evident that machine learning and AI have had a transformative impact on robotics, enabling researchers and practitioners to train robots to perform complex skills without the need for significant user expertise. Indeed, these advances have permeated a wide range of domains within robotics, such as navigation, locomotion, and manipulation. However, learning-based approaches still face significant challenges that prohibit widespread adoption and robust performance. These challenges include i) learning skills from limited data and compute resources, ii) reliably generalizing beyond training distributions, and iii) increasing transparency and predictability of learned policies. This workshop is designed to discuss and explore how structured robot learning can help address and potentially overcome these challenges. Structured robot learning involves incorporating structures from related fields in mathematics and physics into the learning pipeline in order to address these challenges.

Our workshop will specifically focus on improving robot learning by incorporating various forms of structures: dynamical systems, geometry, physics-based, infinite-dimensional feature spaces, as well as logic and symbolic. Indeed, many of these structures have been integral parts of robotics from its infancy. However, we are just starting to see the benefits of incorporating these structures into robot learning. In contrast to typical robot learning approaches, structured learning techniques are proving to be much more sample and computationally efficient, and tend to produce robust and inspectable learned behaviors, across a wide range of applications (e.g., manipulation, locomotion, navigation, and human-robot interaction). These approaches achieve these benefits by injecting suitable inductive biases into the learning framework in form of policy architectures, state and action representations, and parametric and model-level constraints. These approaches attempt to achieve the best of both worlds by combining the unique benefits of analytical approaches and learning-based approaches, leveraging one to address the limitations of the other.

Despite promising results, there is much that is unknown when it comes to structured approaches to robot learning. While scaling (compute and data) has emerged as the dominant paradigm in ML and AI communities, roboticists seem to be divided on whether scaling is the way to reliable and robust robot intelligence. Indeed, structured learning could potentially produce such capable robots without requiring significant compute and data requirements. However, choosing the right structures that simultaneously improve efficiency and generalize across a wide range of variations continues to be challenging. This workshop will bring together senior and junior researchers at the forefront of structured robot learning and related fields to discuss key challenges.

We have designed the workshop to have the following key components to ensure ample high-quality interactions among all participants:

• Tutorial on structured robot learning: It is important that we first orient the workshop participants towards the principled problem of robot learning, the role that structure can play in robot learning, and the sources of structure which will be discussed in this workshop. We will explore several diverse notions of structure seen in the recent literature: dynamical or  physics-based structure, geometric or topological structure, structure from human priors, and structure from feature design. 

• Invited talks: Each invited speaker will focus on their specific research topic related to structured learning and emphasize their latest advances. We already have 7 confirmed diverse speakers from around the world at various professional stages that will provide insights into various notions of structured learning. 

• Discussion/Panel session: An open discussion panel with select speakers will be used to encourage discourse amongst the attendees. Targeted topic questions will be drafted before and during the workshop based on the speaker's research topic and an open questions and discussion session will be encouraged by the organizers.  

• Spotlight presentations: Select workshop paper submissions will be chosen by the organizing committee to be spotlight presentations. The spotlight presentation will provide a platform to highlight the work of  junior researchers amongst senior colleagues. 

• Poster Sessions: We will encourage participants who have contributed workshop papers to exchange ideas through two poster sessions during each coffee break. This will encourage interactions between senior and junior researchers and potentially foster new collaborations and new directions. Submitted papers and participants will be posted on the workshop website.