Best Paper Award 2025 winner interview

Title of the award-winning paper

Port-Hamiltonian Neural ODE Networks on Lie Groups for Robot Dynamics Learning and Control

List of authors

Thai Duong, Abdullah Altawaitan, Jason Stanley, Nikolay Atanasov

Can you briefly introduce yourself and your research background? (each author)

[Thai Duong] I am a postdoctoral researcher at Rice University, Houston, Texas. Last year, I finished my Ph.D. degree in Electrical and Computer Engineering at the University of California San Diego. I obtained an M.S. degree from Oregon State University and a B.S. degree from Hanoi University of Science and Technology, Vietnam. My research interests include robotics, machine learning, control theory and optimization. I have been working on efficient robot learning for planning and control of different robot platforms such as mobile and legged robots, and manipulators.
Personal Website   LinkedIn   Google Scholar

[Abdullah Altawaitan] I am a Ph.D. candidate in Electrical and Computer Engineering at the University of California San Diego. I received both the B.S. and M.S. degrees in Electrical Engineering from Arizona State University. I am also affiliated with the Electrical Engineering Department at Kuwait University, Kuwait. My research interests include machine learning, control theory, and their applications to robotics.
Personal Website   Linkedin   Google Scholar

[Jason Stanley] I am an incoming Ph.D. student in the Electrical and Computer Engineering Department at the University of California, San Diego, where I have also received my B.S. and M.S. Degrees. My research areas include robotics, machine learning, and optimization. I plan to focus my Ph.D. on broadening tools for robotic exploration, including developing new mapping techniques that enable more adaptive and intelligent robot behavior.
Personal Website   LinkedIn   Google Scholar

[Nikolay Atanasov] I am an Associate Professor in the Department of Electrical and Computer Engineering at the University of California San Diego. My research focuses on robotics, control theory, and machine learning with emphasis on active perception problems for autonomous mobile robots. I work on simultaneous localization and mapping (SLAM), motion planning and control techniques for robot navigation in unknown environments, as well as some reinforcement learning for robot control. Prior to joining UCSD’s faculty, I had the privilege of doing my PhD and postdoctoral studies at GRASP lab at the University of Pennsylvania.
Personal Website   LinkedIn   Google Scholar

What is your award-winning paper about in simple terms?

As data is usually limited and costly in robotics, our paper allows learning robot dynamics models accurately and efficiently from just a few minutes of data collection on the robot. Given the learned model, our method can also construct a model-based control policy for stabilization or trajectory tracking. Intuitively, the key idea is to encode what we already know about the robot system, such as universal physics laws and kinematic constraints, in the dynamics model and only learn the rest from data. Thanks to the encoded knowledge, our approach learns faster, with accurate long-term predictions and control performance.

What motivated you to pursue this particular research topic? (Inspiration, practical challenges, gaps in the literature, etc.)

Traditional system identification approaches often use a simplified physics model of the robot dynamics and optimize over a small set of parameters, potentially leading to bias and model errors. Meanwhile, machine learning approaches, which have gained a lot of attention in robotics, are flexible for dynamics modelling due to over-parameterization but unfortunately very data-hungry. This gap motivated us to find a way to obtain the best of both worlds for accurate and efficient robot dynamics learning and control. In 2018, the neural ordinary differential equation (ODE) paper by Chen et al., which allows training a continuous-time model from discrete-time data, was published in NeurIPS. As robot dynamics are commonly described by an ODE via Lagrangian or Hamiltonian formulations, we saw this paper as a bridge connecting physics and machine learning. This inspired a new research direction on dynamics learning and control that we have been pursuing. We were also curious about deploying machine learning models directly on robot systems with embedded computers and this research offered a great way to explore this direction.

What do you consider the key contribution(s) of your paper? (Highlight technical novelty, theoretical advances, practical implications, etc.)

Our key contribution is to impose physics laws and state constraints on the neural network architecture for system dynamics learning. As many robots have states evolving on Lie groups and dynamics satisfying energy conservation principles, we formulated a Hamiltonian-based neural ODE network on a Lie group to approximate the robot dynamics. The encoded structure allows efficient learning from just a handful of state-control trajectories while providing long-term predictions, with guarantees for energy conservation and state constraints by design. Also, this structure enabled an energy-based control design using the learned Hamiltonian dynamics, leading to a unified approach for stabilization and trajectory tracking with various robot platforms. Essentially the same method can be applied with minimal changes to learn dynamics and control various rigid-body robots.

Were there any unexpected challenges or surprises during the research?

One notable challenge we faced was the noise in the real dataset of state-control trajectories, where the state includes the robot’s pose and velocity. While everything worked perfectly in simulation, the real data we collected had considerable amounts of noise in the linear and angular velocity from state estimation. As a result, our Hamiltonian neural ODE network struggled to converge and tended to overfit to the noise. Our solution was to preprocess the data by filtering out the high-frequency noise using a low-pass filter and then adding an L1 norm on the neural network output to the loss function to encourage sparsity in the model and, hence, alleviate overfitting. With this technique, we successfully trained our dynamics model from a small dataset and verified the control design on a real quadrotor.

How do you see this work impacting future research or applications in your field?

In light of the recent ICRA debate on model-based and learning-based approaches for robot automation, our work represents a hybrid approach that combines traditional system modeling and machine learning for system identification and control. We hope that it can spark a lot more interest in integrating model-based and learning-based methods in a principled manner. We think that this is a promising future research direction that can equip autonomous robot systems with online and lifelong learning capabilities, fast adaptation to the world, while still providing guarantees, such as safety and stability, thanks to the preservation of prior knowledge and domain expertise in the model. Also, it would be exciting to apply this technique to articulated robot systems such as quadrupeds and humanoid robots. Our recent preliminary work, presented at ICRA’25, shows that we can improve quadruped jumping performance by learning residual dynamics from limited data.

What are your next steps or future directions following this work?

There are multiple future directions that stem from our work. Instead of learning the system model offline, we can learn and update the model online as the robot dynamics might change over time due to different sensors, motors or onboard computers or disturbances from the environment. Potentially, this can be combined with adaptive control techniques to achieve better stabilization and tracking performance. Another direction is unsupervised model learning, where we can use sensor data such as images or laser scans to infer the robot dynamics, instead of relying on state estimation as labels in the dataset. Our approach can also be extended to multi-robot systems with distributed dynamics learning and control, where each agent infers other robots’ dynamics and generates its control based on local information. Furthermore, learning dynamics and control of multi-rigid body or soft robot systems is very interesting as it requires integrating a kinematic chain, e.g., from a URDF description of the robot, in the neural network architecture. Learning contact dynamics to model robot interaction with the environment is challenging due to its discontinuity but it is a very exciting problem to pursue. Besides continuous-time dynamics, preserving physics structure in a neural discrete-time dynamics model, e.g., via variational integration, will be useful for reinforcement learning and model predictive control techniques. We are very excited about all these new directions and looking forward to seeing new advances in this area.