This paper presents a novel approach to stabilizing periodic orbits in underactuated robots using spiking neuromorphic control. By combining transverse linearization with a greedy Lyapunov-based spike-triggering strategy, the authors provide a theoretical framework and experimental validation for energy-efficient, pulse-based orbital stabilization.
TL;DR
The authors extend a greedy Lyapunov-based spiking control strategy to stabilize periodic orbits in underactuated robots. By linearizing the robot's dynamics around a target trajectory, they transform the problem into a hybrid linear time-varying system and demonstrate practical stability through experiments on a physical butterfly robot.
What Problem Does This Paper Address?
Underactuated robots—systems with fewer actuators than degrees of freedom—are a classic challenge in robotics. Tasks like walking, swimming, or maintaining an oscillation require stabilizing a periodic orbit. Traditionally, this is done with continuous control signals.
However, there is growing interest in neuromorphic control, which uses discrete, event-driven "spikes" (pulses) similar to biological neurons. While neuromorphic systems are potentially more energy-efficient and robust, they are difficult to analyze mathematically. Specifically, it hasn't been clear how to design a spiking controller that guarantees a robot will stay on a complex, nonlinear periodic path, especially when the control pulses have a fixed amplitude.
Methods
The researchers tackle this by breaking the problem into several steps:
- Transverse Linearization: They use a coordinate transformation to describe the robot's state in terms of its progress along the orbit ($\tau$) and its "transverse" error ($\xi$). This reduces the complex nonlinear dynamics to a Linear Time-Varying (LTV) system that is periodic.
- Hybrid System Modeling: Because spikes cause instantaneous jumps in velocity, the system is modeled as a hybrid dynamical system consisting of a continuous "flow" phase and a discrete "jump" phase.
- Greedy Lyapunov-Based Control: They construct a quadratic Lyapunov function $V(\tau, \xi)$—essentially a mathematical measure of "energy" or error. A control spike is triggered only when it is "greedy," meaning the pulse is guaranteed to instantaneously decrease the value of $V$.
- Periodic Riccati Equation: To find the optimal parameters for this Lyapunov function, they solve a periodic version of the Riccati differential equation, ensuring the controller can handle the time-varying nature of the orbit.
Datasets
The paper does not use a standard machine learning dataset. Instead, the approach is evaluated using:
- Numerical Simulations: To solve the periodic Riccati equations and compute reference trajectories.
- Physical Testbed: A "butterfly robot," which consists of a figure-eight-shaped link and a ball that rolls along its surface. This is a standard benchmark for underactuated control.
Key Insights and Ideas
- Coordinate Choice is Crucial: The authors restricted the projection map ($\pi$) to depend only on generalized coordinates (positions). This ensures that the "progress" variable $\tau$ remains continuous even when a control spike causes a sudden jump in velocity, simplifying the stability analysis.
- From LTI to LTV: The "greedy" spiking strategy was previously only well-defined for Linear Time-Invariant (LTI) systems. This work proves it can be extended to periodic LTV systems, which are necessary for orbital stabilization.
- Practical Stability: Since the spikes have a fixed amplitude, the error cannot always be driven to exactly zero. The authors formally define and prove practical orbital stability, meaning the robot's state will converge to and stay within a small, bounded "tube" around the desired orbit.
Key Results
The proposed controller was validated on a physical butterfly robot. The goal was to keep a ball rolling in a stable periodic motion on the robot's frame.
- Robustness: The controller maintained stability across various spike amplitudes ($a \in \{5, 6, 10, 12, 20\} \times 10^{-3}$).
- Convergence: As shown in Figure 3, the transverse variables (errors) successfully converge to a neighborhood of zero. The control signal $w$ consists of dense bursts of spikes that effectively emulate a continuous stabilizing force.
- Tracking Accuracy: Figure 4 demonstrates that the measured joint angles and velocities closely follow the reference trajectories despite real-world disturbances like unmodeled friction, measurement noise, and delays.
Limitations and Future Work
- Local Stability: The current proof and method assume the robot starts relatively close to the target orbit. A "global" stabilization strategy that can catch the robot from any initial state remains a challenge.
- Single Input: The experimental validation focused on a system with one control input ($m=1$). While the authors state the theory can be extended to multiple inputs, they have not yet demonstrated this experimentally.
- Nonlinear Analysis: The stability proof relies on the linearized version of the dynamics. A more rigorous analysis of the full nonlinear hybrid system is left for future work.
Summary updated