Background. Towards the end of robotic systems capable of ubiquitous, persistent deployment in the wild,
researchers across diverse subdisciplines of our field
have exploited the inherent symmetries present in robotic systems and their environment to achieve
drastic improvements in performance, efficiency, and robustness. For example:
- Equivariant architectures for data-driven perception have demonstrated impressive generalization and sample efficiency, all the while reducing model complexity and guaranteeing by design that extraneous transformations will not degrade their predictions.
- A symmetry-aware approach to the filter design has yielded improved convergence properties in state estimation, providing both formal certificates and astonishing empirical accuracy in field deployments.
- As the complexity of both individual agents and multiagent teams has grown, the exploitation of symmetry in control has tamed unwieldy high-dimensional models, sated the appetite of data-hungry methods like reinforcement learning, and aided in the decentralized coordination of large robot swarms.
- introduce new members of the broader robotics community to symmetry-informed methods,
- identify new opportunities to leverage these cross-cutting concepts and apply geometric expertise in new areas,
- align with the IROS 2024 theme "Robotics for Sustainable Development" by reducing the environmental impact of autonomous systems via reduced model complexity, more efficient algorithms, and greater reliability, and
- ultimately bring us closer to the availability of ubiquitous, intelligent robotic systems prepared to tackle society's greatest challenges and play an active role in our daily lives.
Keynote Speakers
Schedule
| Opening Remarks | |
|
Keynote: Chien Erh Lin and Tzu-Yuan Lin, on behalf of Maani Ghaffari
"Computational Symmetry and Learning for Robotics"
Forthcoming mobile robots require efficient generalizable algorithms to operate in challenging and unknown environments without human intervention while collaborating with humans. Today, despite the rapid progress in robotics and autonomy, no robot can deliver human-level performance in everyday tasks and missions such as search and rescue, exploration, and environmental monitoring and conservation. In this talk, I will put forward a vision for enabling efficiency and generalization requirements of real-world robotics via computational symmetry and learning. I will walk you through structures that arise from combining symmetry, geometry, and learning in various foundational problems in robotics and showcase their performance in experiments ranging from perception to control. In the end, I will share my thoughts on promising future directions and opportunities based on lessons learned on the field and campus.
|
|
|
Keynote: Robert Mahony
"Galilean Space-Time in Robotics"
Rigid-body or pose symmetry lies at the foundation of the
spatial geometry that underlies most of classical robotics. Recent
work has shown that there is an "extended pose" symmetry that has
been demonstrated to provide significant benefits in inertial
navigation problems. In this talk I show that the extended pose
symmetry is itself a specialisation of the natural symmetry
associated with Galilean space-time and show how this formalisation
leads to a powerful modelling framework for dynamic robotic
systems.
|
|
|
Keynote: Robin Walters
"Pushing the Limits of Equivariant Neural Networks"
Incorporating symmetry constraints into neural networks has resulted in models called equivariant neural networks (ENNs) which are more sample efficient and generalize better. This has proved to be a crucial benefit in domains such as robotic manipulation where on-robot reinforcement learning can be very time consuming and expert demonstrations are expensive. I will discuss several applications of ENNs such as vehicle trajectory prediction, fluid dynamics modeling, robotic manipulation, world modeling, and differentiable planning. However, there are also limits to the effectiveness of current ENNs. In many applications the observation type makes the symmetry in the domain hard to exploit. I will discuss solutions to this problem such as relaxed equivariance and models which take advantage of latent symmetry.
|
|
|
Poster Session + Coffee
This session will take place at the foyer of the breakout rooms (Halls 2,3,4) on boards 7.1 - 7.8.
|
|
|
Keynote: Elise van der Pol (virtual speaker)
"Symmetry & Structure in RL - State of the art, challenges, and future directions"
In this talk, I will give an overview of the state of the field of symmetries and structure in reinforcement learning. I'll discuss the why and how of symmetric RL, standing challenges in the field, what work has been done to address those challenges and identify key directions for future work.
|
|
|
Keynote: Amanda Prorok (virtual speaker)
"Leveraging Symmetry for Modeling Multi-Agent Interaction"
(Abstract forthcoming.)
|
|
|
Panel Discussion
Our keynote speakers will join us once more to discuss key takeaways and open questions.
|
|
| Award Session + Closing Remarks |
All times are UTC+4. Schedule will be finalized closer to the workshop.
Call for Papers
We invite contributions exploring the role of symmetry across diverse problems in robotics and autonomy, including (but not limited to):
- Geometric mechanics and symmetry in locomotion planning and control
- Conservation laws and motion planning for nonholonomic vehicles
- Formal certificates or experimental verification of equivariant filters
- Approximately equivariant architectures for working with broken symmetry
- Feedback linearization and differential flatness in the presence of symmetry
- Model order reduction via symmetry
- Equivariant neural representations
- Symmetry as a prior in physics-informed machine learning
- Discrete symmetries in biological and robotic systems
- Equivariant deep learning
- Symmetries in geometric perception (3D reconstruction, registration, 3D object detection)
- Symmetries with proprioceptive sensors (IMU integration etc.)
- Symmetry-informed optimizers
- Sample complexity and robustness benefits in equivariant machine learning
- Symmetries in multiagent systems (flocking, tracking, coordination)
- Graph neural networks for decentralized multiagent autonomy
Author Guidelines
- We welcome the contribution of short papers / extended abstracts of 2-4 pages in 2-column IEEE conference format (including all figures and appendices but excluding references), to give a chance to authors of already published or ongoing works to present their work at the workshop.
- The workshop is non-archival (i.e. contribution should not prohibit submission to other venues), and preliminary or late-breaking results are welcome. Already-published works should mention where the work has previously been published. Contributions will be reviewed (single blind) for basic quality and relevance to the workshop.
- Accepted abstracts will be available on the workshop website and presented in poster format during the workshop.
- Extended abstracts should be submitted via EasyChair: https://easychair.org/conferences/?conf=symrob2024
Important Dates
All deadlines are AoE (Anywhere on Earth).
- Paper Submission:
August 16, 2024September 16, 2024 - Acceptance Notification:
August 30, 2024September 23, 2024 - Camera-Ready Submission: October 11, 2024
- Workshop: October 14, 2024
Best Poster Award
To encourage contributions, we will present a Best Poster Award to one of the contributions (to be judged by invited speakers and perhaps other senior researchers), with a monetary prize supported by the IEEE RAS TC's on "Algorithms for Planning and Control of Robot Motion" and "Computer & Robot Vision".
Accepted Papers
The final extended abstracts for all accepted contributions will be posted on the website before the workshop,
and authors will present their work during the poster session. The accepted extended abstracts are listed below.
- Equivariant IMU Preintegration with Biases: an Inhomogeneous Galilean Group Approach,
- Hybrid-MSCEqF: An Equivariant Visual-Inertial Navigation System,
- Differentiable E(2)-Equivariant Graph Planning for Navigation,
- Towards Multi-Sensor Equivariant Filter Design,
- Exploring Diverse Quadrupedal Gaits through Symmetry Breaking,
- Leveraging Symmetry in RL-based Legged Locomotion Control,
- Geometric Algebra Grasp Diffusion for Dexterous Manipulators,
- Contact-rich SE(3)-Equivariant Robot Manipulation Task Learning via Geometric Impedance Control,
- Leveraging Symmetry to Accelerate Learning of Trajectory Tracking Controllers for Free-Flying Robotic Systems,
Organizers
Acknowledgements
The organizers gratefully acknowledge the financial support provided for the workshop by the IEEE RAS TC's on
"Algorithms for Planning and Control of Robot Motion" and "Computer & Robot Vision".
We also appreciate the endorsement of the workshop by the IEEE RAS TC's on "Multi-Robot Systems" and "Robot Learning".