Introduction
Accurate dynamics modeling is a cornerstone of robotics, yet traditional first-principles models often fall short in complex, real-world environments. While data-driven approaches like standard Neural Networks offer flexibility, they frequently act as "black boxes" that ignore fundamental physical laws. This leads to issues such as energy violation in long-term predictions and mathematical singularities (e.g., gimbal lock) when representing 3D orientations with Euler angles. This research introduces a structured Neural ODE architecture that embeds Hamiltonian mechanics and $SE(3)$ manifold constraints directly into the learning process, ensuring physically consistent and singularity-free dynamics.
Methodology
The proposed framework leverages Hamiltonian-based Neural ODEs formulated on the Special Euclidean group $SE(3)$. By learning the derivative of the state through a Hamiltonian lens, the model decomposes total energy into kinetic and potential components represented by separate neural networks. Specifically, the architecture approximates the inverse mass matrix ($M^{-1}$), potential energy ($V$), and the control input matrix ($g$).
An important insight of this approach is its scale invariance; the model learns a scaled version of the true physics (factor $eta$), which is sufficient for control as relative ratios remain consistent. This structured approach allows for the application of Interconnection and Damping Assignment Passivity-Based Control (IDA-PBC). By "shaping" the learned energy surface, the system achieves stable trajectory tracking without needing explicit knowledge of the true physical parameters.
Experimental Evaluation and Results
The model was evaluated using simulated datasets for a pendulum, a fully-actuated rigid body, and a Crazyflie quadrotor in the PyBullet physics engine. The Hamiltonian-based Neural ODE demonstrated superior long-term stability and physical fidelity compared to unstructured models.
Key performance highlights include:
- Energy Conservation: In long-term rollouts without control, the model maintained near-perfect energy conservation, whereas standard models typically drift.
- Geometric Accuracy: The model perfectly maintained $SE(3)$ constraints, such as the orthogonality of rotation matrices, with errors staying near machine epsilon ($10^{-7}$).
- Control Performance: The learned model enabled a quadrotor to track complex trajectories with a real-time computation frequency of approximately 3.5ms per step.
Conclusion
By integrating Hamiltonian mechanics and manifold geometry into Neural ODEs, this research provides a robust method for learning robot dynamics that are physically guaranteed to be valid. While current work focuses on conservative systems, future extensions to include dissipation and multi-body dynamics will further enhance the applicability of this structured learning approach to real-world robotic systems.
Summary updated