Hamiltonian Dynamics Learning from Point Cloud Observations for Nonholonomic Mobile Robot Control

Abdullah Altawaitan

Jason Stanley

Sambaran Ghosal

Thai Duong

Nikolay Atanasov

Department of Electrical and Computer Engineering
University of California, San Diego
ICRA, 2024.

[Paper]

[Code]

Reliable autonomous robot navigation requires adapting the control policy to dynamics changes under different operational conditions. While hand-designed dynamics models from first principles with few parameters may not be able to capture the dynamics variations, data-driven approaches requires a large amount of training data and perfect state estimation. Hence, they are prone to state estimation errors or dependent on a motion capture system, limiting their applications in the real world. In this paper, we aim to learn the robot dynamics directly from sensor observations to avert state estimation errors, while embedding a Hamiltonian structure in the machine learning model to improve data efficiency, generalization and facilitate control design. The consistency between two consecutive observations, e.g. appearance of the same objects, corners or edges, forms a loss function, which in turns, is used to udpate the dynamics model via gradient descent. We develop an energy-shaping model-based controller of rigid-body robots for trajectory tracking with the learned dynamics. We verify our approach in simulated and real experiments with nonholonomic wheeled robots.

Paper

Abdullah Altawaitan, Jason Stanley, Sambaran Ghosal, Thai Duong, Nikolay Atanasov

Hamiltonian Dynamics Learning from Point Cloud Observations for Nonholonomic Mobile Robot Control

Submitted to ICRA, 2024.

[pdf]

Video

Acknowledgements

We gratefully acknowledge support from NSF CCF-2112665 (TILOS).
This webpage template was borrowed from https://akanazawa.github.io/cmr/.