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

TL;DR

This work addresses trajectory tracking for mobile robots with uncertain dynamics by learning a dynamics model directly from point-cloud observations. It employs a Hamiltonian neural ODE (HNODE) with a cycle-consistency loss that aligns predicted motion from the learned dynamics with point-cloud registrations, enabling training without explicit state estimation. The learned port-Hamiltonian model ($\mathbf{M}_{\theta}, \mathcal{H}_{\theta}, \mathbf{D}_{\theta}, \mathbf{g}_{\theta}$) supports an energy-shaping control law via IDA-PBC, producing a control input $\mathbf{u} = \mathbf{u}_{ES} + \mathbf{u}_{DI}$ that drives the system toward a desired Hamiltonian $\mathcal{H}_d$. The approach is validated on real nonholonomic wheeled robots (and in simulation) showing accurate dynamics learning, energy-conserving behavior, and improved trajectory tracking compared to nominal models. This has practical impact for robust autonomous navigation in changing conditions without relying on full state estimation pipelines.

Abstract

Reliable autonomous navigation requires adapting the control policy of a mobile robot in response to dynamics changes in different operational conditions. Hand-designed dynamics models may struggle to capture model variations due to a limited set of parameters. Data-driven dynamics learning approaches offer higher model capacity and better generalization but require large amounts of state-labeled data. This paper develops an approach for learning robot dynamics directly from point-cloud observations, removing the need and associated errors of state estimation, while embedding Hamiltonian structure in the dynamics model to improve data efficiency. We design an observation-space loss that relates motion prediction from the dynamics model with motion prediction from point-cloud registration to train a Hamiltonian neural ordinary differential equation. The learned Hamiltonian model enables the design of an energy-shaping model-based tracking controller for rigid-body robots. We demonstrate dynamics learning and tracking control on a real nonholonomic wheeled robot.

Figures (5)

• Figure 1: Clearpath Jackal robot equipped with a LiDAR.
• Figure 2: Illustration of our observation-space loss function design.
• Figure 3: Evaluation of our Hamiltonian Neural ODE network (a)-(g) along a trajectory and (h) pose stabilization in the Sapien simulator.
• Figure 4: Evaluation of our Hamiltonian neural ODE network (a)-(g) along a trajectory and (h) pose stabilization on a real Jackal.
• Figure 5: Evaluating nominal and learned model performance in tracking circular and lemniscate trajectories.