Introduction

Most robotic applications, including manipulation, navigation and locomotion, require our robots to interact with the physical world, which is rich in geometric and algebraic structures. This workshop is intended to provide a forum to discuss whether and how similar geometric and algebraic structures can benefit various aspects of robot learning, including perception, planning, control, and beyond.

Robot learning researchers have exploited different forms of geometric and algebraic structure to improve efficiency, generalization capability, robustness, scalability, etc. for different aspects of robot learning. In perception and representation learning, researchers exploit the geometric and algebraic structure of disentangled representation, factored or symbolic features, and compositional scene understanding. In decision-making, these structures have been explored in reinforcement learning and planning algorithms through geometric deep learning, skill learning, goal-conditioned learning, and language compositionality. The goal of this workshop is to bring together researchers who work on different aspects of robot learning, identify common tools and methodologies for incorporating geometric and algebraic structures for robot learning, and potentially foster new ideas on leveraging such structures.

The intended audience is primarily researchers in robot learning, especially those who are interested in enhancing learning efficiency, generalization capability, robustness, scalability, etc. of pure data-driven approaches. The intended audience also includes (non-learning) roboticists and theory researchers who are knowledgeable in areas such as group theory, representation theory, category theory, symbolic reasoning, etc., and would like to explore robot learning as a potential application scenario. 

Discussion topics

• What robotics applications benefit from having geometric and algebraic structures, such as manipulation, navigation, locomotion, etc?

• How to build practical learning algorithms that can leverage structures and be applied to real-world physical robots? 

• What are the scenarios when geometric and algebraic structures are not helpful for robot learning, and how could we learn from these failure cases?

• How can geometric and algebraic structures, such as spatial (e.g., objects), temporal (e.g., skills), or parallel (e.g., multi-objectives) compositionality, be formulated with mathematical tools?

• What types of geometric and algebraic structures exist in robot learning? Are there any symmetries worth considering beyond Euclidean symmetry (translation, rotation, reflection) and permutation symmetry, such as scale equivariance?

• What practical benefits do they bring to robotics, especially robot learning, such as generalizability, efficiency, and scalability? How can they aid the robotics stack, such as perception, planning, and control?

• How do symmetry and compositionality interact? For example, to grasp multiple objects, the algorithm needs to consider both object symmetry and object interchangeability (compositionality).

Invited Speakers

Tentative Schedule

Example directions

• 1. Hierarchical, Modular, and Compositional Structures:

• Compositional Models in Robot Perception, Planning, and Learning: Development of models that interpret or construct complex robotic behaviors from simpler components, facilitating advanced perception and decision-making capabilities.

• Hierarchical Reinforcement Learning: Employing hierarchical structures to decompose complex tasks into simpler, manageable sub-tasks, enhancing learning efficiency and practical applicability.

• Compositional Understanding of Objects: Developing approaches that allow robots to recognize and manipulate objects by understanding their compositional properties and how they relate to each other in various configurations.

• Enhancing Scalability and Generalizability in Robotic Tasks: Investigating systems that leverage modular and hierarchical designs to improve the scalability and generalizability across different robotic applications.

• 2. Symmetry in Robot Learning:

• Symmetry in Perception, Planning, and Control: Exploring the role of symmetry in simplifying and enhancing the reliability of robotic control systems across different environments.

• Equivariant Neural Networks: Utilizing neural networks that maintain the same output under certain transformations to ensure consistent performance across varied perceptions and actions.

• Group Theory for Robotics Applications: Applying principles of group theory to create robust and efficient algorithms that exploit symmetrical properties in robot learning tasks.

• 3. Semantic-Geometric Structures:

• Integration of Semantic Information and Geometric Data: Examining how robots can process structured semantic inputs from language and integrate them with geometric information from their environment to enhance task understanding and execution.

• Language-Conditioned Robot Learning: Investigating how compositional and algebraic structures in language can be leveraged to improve the interpretability and effectiveness of robot learning models.

• Leveraging Algebraic Structures in Language for Robotic Understanding: Exploring how algebraic concepts inherent in language can be used to inform and enhance robotic learning and interaction strategies.

Organizers

Contact

Please feel free to reach out to the organizer group email (gas-rss2024@googlegroups.com) or Linfeng Zhao (zlf0625@gmail.com) for any questions.