Description

Robotics researchers have often leveraged geometric or physical structure in the world to make challenging planning and control problems easier. Recent work has applied this idea to robotic learning in order to improve sample efficiency and generalization. One approach is to use geometric deep learning by encoding symmetries in the structure of neural network model architectures. Geometric deep learning has had striking successes in molecular biology, computer vision, particle physics, and elsewhere. The question is how geometric deep learning methods can be employed in robotics.

The goal of the Symmetries in Robot Learning workshop is to bring together robotics researchers with researchers from the geometric deep learning community to discuss this question and shape the vision of the field of symmetry in robot learning. We will encourage researchers with experience in using symmetric learning in robotics to share ideas about how to leverage problem symmetries. At the same time, we will encourage researchers from the equivariant machine learning community to share their experiences in other fields that might inspire new approaches to leveraging structure in robotics.

09:00 am–09:05 am: Opening remarks

09:05 am–09:35 am: Invited talk, Tess Smidt

Intuition for Data Types and Interactions in Euclidean Symmetry-Equivariant Neural Networks

09:35 am–10:05 am: Invited talk, Hao Su

Learning Generalizable Policies by 3D Reinforcement Learning

10:05 am–10:25 am: Oral presentation 1

Robotic Manipulation Learning with Equivariant Descriptor Fields: Generative Modeling, Bi-equivariance, Steerability, and Locality

10:25 am–11:10 am: Coffee break + Poster session

11:10 am–11:40 am: Invited talk, Maani Ghaffari

A Tour of Recent Symmetry-Preserving Methods for Mobile Robots

11:40 am–12:00 pm: Panel discussion (Tess Smidt, Hao Su, Maani Ghaffari)

12:00 pm–01:30 pm: Lunch break

01:30 pm–02:00 pm: Invited talk, Andy Zeng

From the Symmetries in Words to Actions

02:00 pm–02:50 pm: Contributed paper presentations (5 mins each)

1. Morphological symmetries in robot learning

2. Edge Grasp Network: A Graph-Based SE(3)-invariant Approach to Grasp Detection

3. Euclidean Equivariant Models for Generative Graphical Inverse Kinematics

4. Point-based Correspondence Estimation for Cloth Alignment and Manipulation

5. Continual Reinforcement Learning with Group Symmetries

6. Geometric Regularity with Robot Intrinsic Symmetry in Reinforcement Learning

7. Geometric Algebra Transformers

8. Spatial Generalization of Visual Imitation Learning with Position-Invariant Regularization

02:50 pm–03:40 pm: Coffee break + Poster session

03:40 pm–04:00 pm: Oral presentation 2

Progressive Learning for Physics-informed Neural Motion Planning

04:00 pm–04:30 pm: Invited talk, Erik Bekkers

Group Equivariant Deep Learning in Convolutional Form

04:30 pm–05:00 pm: Invited talk, He Wang

Title TBD

05:00 pm–05:20 pm: Panel discussion (Andy Zeng, Erik Bekkers, He Wang)

05:20 pm–05:30 pm: Closing remarks

• June 9th, 2023  (AoE): Submission Deadline

• June 23rd, 2023: Acceptance Notification

• July 2nd, 2023: Camera Ready Deadline

• July 10th, 2023: Workshop