Introduction
Robotic air hockey has traditionally been the domain of stationary robotic arms. However, researchers at EPFL have introduced a more dynamic alternative: a free-floating, propeller-actuated hovercraft designed to act as an "autonomous puck." While this approach offers greater mobility, it introduces significant complexity in system identification. Mapping control signals for multiple brushless motors to the precise movement of a lightweight foam disc involves navigating non-linear thrust curves, motor response delays, and physical asymmetries. This paper presents a method for identifying these complex dynamics by combining first-principles physics with end-to-end deep learning.
Hardware and Experimental Setup
The experimental platform is a custom-built 14cm hovercraft. It features a 3D-printed ring housing six racing drone propellers, powered by an ESP32 microcontroller. To estimate local velocity, the vehicle utilizes an onboard optical mouse sensor, while ground truth position and orientation are provided by an external OptiTrack motion capture system operating at 240Hz. The entire assembly weighs approximately 137g, making it highly sensitive to even minor aerodynamic and mechanical fluctuations.
Physics-Inspired System Identification
Rather than treating the hovercraft as a "black box," the authors employ a physics-inspired learning approach. They embed the equations of 2D rigid body dynamics directly into the learning loop, parameterizing specific physical constants that are difficult to measure manually.
Modeling Dynamics
The system models thrust as a third-order polynomial and accounts for the latency in motor response as a first-order system with a learnable time constant ($\tau$). By using a batched prediction loss—a technique common in training recurrent neural networks—the model parameters are optimized using the Adam optimizer in PyTorch.
Physical Regularization
A key innovation in this approach is the use of physical regularization. To prevent the model from learning mathematically valid but physically impossible behaviors, the researchers introduced a monotonicity constraint. This ensures that the learned thrust model reflects reality: an increase in power must always result in an increase in thrust. Without this constraint, the deep learning model tended to produce nonsensical thrust profiles that failed in real-world applications.
Control Architecture
The control system uses a hierarchical architecture to manage the identified non-linearities:
- High-Level Control: A Linear Quadratic Regulator (LQR) determines the desired accelerations required to follow a trajectory.
- Low-Level Mapping: These accelerations are mapped to specific motor voltages by solving a Quadratic Program (QP) in real-time using the PIQP solver. This allows the system to handle motor saturation and non-linear thrust curves efficiently at frequencies exceeding 100 Hz.
Results and Discussion
The researchers evaluated five model variants (M0 through M4) with increasing levels of complexity. The most sophisticated model, M4, achieved a Root Mean Squared Error (RMSE) of 10.8 mm in position and 0.387 rad in orientation over a 0.2-second prediction horizon.
Key Insights
- Motor Dynamics vs. Inertia: Interestingly, the study found that modeling the first-order motor dynamics (the delay in propeller spin-up) was more critical for prediction accuracy than the precise identification of the vehicle's moment of inertia.
- Interpretability: By using a first-principles model instead of a pure configuration matrix, the researchers maintained interpretability. They could visualize actual thrust values and physical parameters, which is often impossible with purely connectionist deep learning models.
- Closed-Loop Performance: In tracking tests, the M4 model successfully eliminated the steady-state offsets and sluggish response times observed in simpler models that ignored motor dynamics.
Conclusion and Future Work
This research demonstrates that end-to-end deep learning, when constrained by physical principles, can effectively identify the dynamics of complex robotic systems. While the current LQR-based control is effective, it does not explicitly account for state and input constraints.
Future work will focus on implementing Model Predictive Control (MPC) to better manage the physical limits of the motors and the boundaries of the air hockey table. The ultimate objective is to integrate real-time puck tracking and move toward time-optimal control, enabling the hovercraft to compete against human players in high-speed matches.
Summary updated