Abstract
Robotic systems operating in real-world environments frequently encounter complex, non-linear disturbances such as gusty winds or shifting currents. Traditional adaptive control methods often struggle with these forces due to the need for manually defined mathematical structures, while pure machine learning approaches often lack physical consistency and real-time efficiency. This paper presents a hybrid framework that combines physics-informed machine learning with geometric adaptive control. By learning non-linear disturbance features offline using a Hamiltonian Neural Ordinary Differential Equation (ODE) and adapting to them online via geometric tracking errors on the SE(3) manifold, the proposed method enables robots to maintain stability and performance even under severe conditions like rotor failure.
Introduction
Deploying robots in unstructured environments requires controllers that can handle unpredictable external forces. While adaptive control is a standard solution, its effectiveness is often limited by the engineer's ability to pre-specify the functional form of the disturbance. For complex phenomena like aerodynamic ground effects or non-linear friction, such formulas are rarely available.
Deep learning offers a potential solution for modeling these complexities, but training large neural networks in real-time control loops is computationally prohibitive and often ignores the underlying physics. Furthermore, standard models frequently fail to account for the unique geometry of 3D rotation (the SE(3) manifold), leading to mathematical singularities and physical inconsistencies. This work addresses these challenges by decoupling the learning of complex feature shapes from the real-time estimation of their magnitudes.
Methodology
The proposed framework consists of three integrated components designed to ensure both flexibility and physical rigor:
- Offline Feature Learning (Hamiltonian Neural ODE): Instead of a "black-box" network, the authors utilize a Hamiltonian-based Neural ODE. This architecture is constrained to respect Hamilton’s equations of motion, ensuring that learned disturbances are physically plausible. It identifies a set of non-linear "features" ($W_\theta$) representing potential disturbances from offline data.
- Online Geometric Adaptation: During operation, the robot uses a geometric adaptation law to update the weights ($a$) of these pre-learned features in real-time. This law is driven by tracking errors in position, orientation, and momentum, allowing for rapid response to environmental changes.
- Geometric Control on SE(3): The system employs Interconnection and Damping Assignment Passivity-Based Control (IDA-PBC). Formulated directly on the SE(3) manifold, this "energy-shaping" method avoids the glitches associated with Euler angles and ensures stable behavior across the entire rotation group.
Experimental Evaluation
The researchers validated their approach using high-fidelity simulations in PyBullet across two primary scenarios:
- Fully-Actuated Pendulum: Focused on learning and compensating for unknown friction coefficients across various disturbance realizations.
- Under-actuated Quadrotor (Crazyflie): A complex test involving horizontal wind, aerodynamic drag, and a "Scenario 2" failure mode where two rotors lose 20% of their thrust mid-flight.
Key Results and Discussion
The hybrid approach demonstrated significant improvements over both non-adaptive baselines and standard Disturbance Observers (DOB). In quadrotor simulations, the proposed adaptive controller reduced tracking errors by approximately fivefold.
A critical advantage of this method is its reduced "lag" compared to DOBs. Because the model already possesses the "shapes" of potential disturbances from the offline phase, it can identify the magnitude of velocity-dependent changes much faster than an observer that must re-discover the disturbance structure at every time step. This speed was vital in the rotor failure tests, where the robot successfully recovered and continued its mission, while non-adaptive controllers resulted in a crash within seconds.
Conclusion
By splitting the control problem into offline feature learning and online weight adaptation, this work provides a robust path for deploying deep learning in safety-critical robotic applications. The integration of Hamiltonian physics and SE(3) geometry ensures that the learned models remain physically consistent and mathematically sound. Future work will focus on transitioning these results from simulation to physical hardware, addressing real-world challenges such as sensor noise and communication latency.
Summary updated