In the rapidly evolving field of robotics, bionic robots have emerged as a pivotal research direction due to their exceptional adaptability in complex environments, ability to generate diverse gaits, and potential for energy efficiency optimization. As technology progresses, control methods for these bionic robots have diversified, encompassing model-based, optimization-based, behavior-based, learning-based, and bio-inspired approaches. Each method offers distinct advantages tailored to specific applications. This review focuses on Central Pattern Generator (CPG) control, a bio-inspired strategy that mimics neural circuits in vertebrates to produce rhythmic movements without relying on precise system modeling or external signals. CPG’s inherent properties, such as periodicity, stability, coordination, robustness, tunability, and continuity, make it highly suitable for enhancing the locomotion performance of bionic robots in dynamic settings. Here, I explore the mathematical models of CPG, its applications across various bionic robot platforms, and the integration with advanced algorithms, providing a comprehensive perspective to propel future developments in bionic control.
Control methods for bionic robots can be broadly categorized, each with unique core ideas and characteristics. To summarize, I present a comparison in Table 1.
| Control Method | Core Idea | Characteristics |
|---|---|---|
| Model-Based | Dynamics and kinematics modeling | Precise and predictable, suitable for structured scenarios, heavily model-dependent |
| Optimization-Based | Mathematical optimization algorithms for multi-objective goals | Adapts to dynamic environments, high accuracy, computationally complex, requires real-time processing |
| Behavior-Based | Modular perception-action with priority adjustment | Fast response, adaptable, lacks global planning, prone to instability |
| Learning-Based | Reinforcement learning or neural networks for behavior optimization via trial-and-error | Highly adaptive, relies on large datasets, computationally intensive |
| Bio-Inspired | Simulates biological neural networks to generate periodic signals | Biologically efficient, parameters are complex to debug |
Among these, CPG control stands out for its ability to autonomously produce sustained oscillatory signals through mutual inhibition and activation among neurons, eliminating the need for external inputs. The key properties of CPG that enhance bionic robot locomotion are detailed in Table 2.
| Property | Description |
|---|---|
| Periodicity | CPG generates signals with stable cycles to maintain rhythmic motion |
| Stability | CPG can recover to original periodic states under external disturbances |
| Coordination | Oscillators maintain appropriate phase relationships for coordinated movement |
| Robustness | CPG adapts to environmental changes or load variations while maintaining stable output |
| Tunability | Output signals can be dynamically adjusted via oscillator parameters |
| Continuity | Output signals are smooth and continuous, avoiding unnecessary oscillations |
This review delves into the progress of CPG control in bionic robots. First, I introduce common CPG mathematical models, including the Hopf, Matsuoka, and Kimura models. Then, I examine applications in legged bionic robots, humanoid robots, and underwater bionic robots, highlighting challenges and solutions. Furthermore, I analyze the fusion of CPG with optimization algorithms, neural networks, and reinforcement learning. Finally, I offer insights into future directions for CPG control in bionic robot design and applications.
Mathematical Models of CPG
In bionic robotics, researchers employ mathematical modeling to emulate CPG functionality for locomotion control systems. CPG models are abstracted at different levels: oscillator models composed of coupled nonlinear oscillators, and neuron models that simulate action potentials and dynamic behaviors of biological neurons. For practical engineering, widely used models include the Hopf, Matsuoka, and Kimura models. These models enable the generation of rhythmic signals essential for controlling bionic robots.
The Hopf oscillator is a prevalent oscillator model due to its simplicity in controlling output signal amplitude and frequency through parameter adjustments. Its mathematical expression is given by:
$$
\dot{x} = \gamma(\mu – r^2)x – \omega y, \quad \dot{y} = \gamma(\mu – r^2)y + \omega x
$$
where \( x \) and \( y \) are state variables determining harmonic output, \( \omega \) represents the oscillation frequency of the limit cycle, and \( \mu \) controls the amplitude and radius of the limit cycle with \( r = \sqrt{x^2 + y^2} \). This model is effective for generating smooth periodic signals in bionic robot joints.
The Matsuoka oscillator, a typical neuron model, incorporates fatigue characteristics to reflect biological properties. It is defined as:
$$
\tau_u \dot{u}_i = -u_i – \beta v_i + \sum_{j=1}^n w_{ij} y_j + u_{0,i} + \text{feed}_i, \quad \tau_v \dot{v}_i = -v_i + y_i, \quad y_i = \max(0, u_i)
$$
where \( u_i \) and \( v_i \) are the membrane potential and adaptation state of the \( i \)-th neuron, \( \tau_u \) and \( \tau_v \) are time constants for rise and fatigue times, \( w_{ij} \) is the connection weight between neurons, \( u_{0,i} \) is external input, \( \text{feed}_i \) is higher-level controller input, \( \beta \) is the inhibition weight, and \( y_i \) is the output signal. This model supports adaptive control but typically produces only positive signals.
The Kimura model, an improvement on Matsuoka, enables both positive and negative outputs by simulating antagonistic muscle groups. It consists of two mutually inhibiting neurons for flexor and extensor control, expressed as:
$$
\tau_f \dot{u}_f = -u_f – \beta v_f – \gamma y_e + u_{0,f} + \text{feedback}_f, \quad \tau_e \dot{u}_e = -u_e – \beta v_e – \gamma y_f + u_{0,e} + \text{feedback}_e, \quad \tau_a \dot{v}_f = -v_f + y_f, \quad \tau_a \dot{v}_e = -v_e + y_e, \quad y_f = \max(0, u_f), \quad y_e = \max(0, u_e)
$$
where subscripts \( f \) and \( e \) denote flexor and extensor neurons, \( \tau_f \) and \( \tau_e \) are time constants affecting output period, \( u \) and \( v \) are internal and inhibitory states, \( \gamma \) is inter-neuron inhibition coefficient, \( \beta \) is adaptation coefficient, and \( y \) is the oscillator output. This model is suitable for generating biomimetic gaits in bionic robots.
To compare these models, their output waveforms exhibit distinct features: Hopf offers smooth limit cycles, Matsuoka enables multi-joint coordination through neuro-feedback coupling, and Kimura mimics excitation-inhibition dynamics for lifelike stepping. In bionic robot applications, selecting a model depends on the required signal properties and control complexity.
Applications of CPG in Bionic Robots
CPG control has been extensively applied to various bionic robot platforms, enhancing locomotion through rhythmic signal generation. I discuss its use in legged bionic robots, humanoid robots, spinal motion control, and underwater bionic robots, emphasizing how CPG networks facilitate adaptive and coordinated movements.
Legged Bionic Robot Locomotion Control
Legged bionic robots, known for their agility and mobility, benefit from CPG-based control that mimics vertebrate neural structures. By constructing CPG networks with multiple units, diverse gaits such as wave, quadrupedal, tripod, and bipedal can be generated. For instance, a three-neuron CPG model incorporating force feedback has been used to adjust gait rhythm and joint angles in hexapod bionic robots based on foot pressure signals. Similarly, a multi-layered CPG approach with feedback loops improves adaptive walking on slopes for quadruped bionic robots. Another study utilized a bidirectional diffusion-coupled CPG topology with Hopf oscillators to achieve stable gait transitions in hexapod bionic robots, demonstrating robustness in various stepping patterns. The control structure typically includes leg trajectory generators and inverse kinematics modules to precisely actuate joint angles, showcasing CPG’s versatility in legged bionic robot systems.
Humanoid Robot Motion Control
Humanoid bionic robots present higher challenges due to multi-joint coordination, balance, and gait stability. CPG control in this domain often employs joint-space trajectory generation, but complexity arises with many degrees of freedom. To address this, workspace trajectory generation methods have gained prominence, using fewer CPG parameters to simplify networks. For example, Kimura models have generated Cartesian trajectories for foot and center-of-mass motion in biped bionic robots, integrating sensor feedback for adaptive walking. Additionally, self-learning CPG models can acquire arbitrary periodic signals from demonstration trajectories, enhancing flexibility. Asymmetric CPG structures inspired by biological systems also offer insights for stable bipedal locomotion in bionic robots. However, dynamic behavior control in humanoid bionic robots remains challenging, necessitating further research into synchronized multi-limb adjustment mechanisms.
Spinal Motion Control in Bionic Robots
Spinal flexibility is crucial for coordinated motion in bionic robots, reducing energy consumption and improving stability. CPG networks transmit rhythmic signals to spine and limbs, synchronizing body movements. In quadruped bionic robots, parallel spine designs controlled by CPG networks minimize lateral sway during trotting. A hierarchical three-layer CPG structure further refines control by coordinating spine and leg motions for high-speed gaits, optimizing velocity through parameter analysis. Despite progress, highly flexible spines in bionic robots are still nascent, with soft actuation emerging as a key research direction to address control complexities in multi-degree-of-freedom systems.
Underwater Bionic Robot Motion Control
Underwater bionic robots operate in complex fluid dynamics environments where CPG control offers adaptive advantages over traditional methods. For robotic fish, CPG generates periodic signals to mimic natural swimming, enabling efficient propulsion and maneuverability. Phase oscillators have been used to coordinate caudal and anal fin movements, enhancing yaw stability in bionic robotic fish. Modified CPG models allow real-time parameter tuning for multi-modal motion transitions, such as generating triangular or sinusoidal waveforms. Amphibious bionic robots like salamander-inspired platforms utilize CPG networks with coupled oscillators to switch between swimming and walking gaits via midbrain stimulation, demonstrating environmental adaptability. These applications underscore CPG’s potential in overcoming hydrodynamic challenges for underwater bionic robots.
Key Control Technologies for CPG
To optimize CPG performance in bionic robots, advanced technologies have been integrated for parameter modulation and output waveform shaping. These include multi-objective algorithms and intelligent learning methods, which enhance adaptability and efficiency.
Multi-Objective Algorithms for CPG Parameter Modulation
CPG parameters, such as frequency, amplitude, and phase, directly influence locomotion patterns in bionic robots. Tuning these parameters is a multi-objective optimization problem balancing stability, energy efficiency, speed, and coordination. Algorithms like Genetic Algorithm (GA), Particle Swarm Optimization (PSO), and Non-dominated Sorting Genetic Algorithm II (NSGA-II) are employed to find optimal parameter sets. For instance, a PSO framework optimized amplitude and frequency in a bionic robotic fish CPG network, increasing swimming speed by 6.7% and propulsion efficiency by 11.67%. This approach allows bionic robots to achieve superior locomotion performance through automated parameter search, reducing manual tuning efforts.
Intelligent Algorithms for CPG Output Waveform Modulation
Traditional CPG models often produce simple sinusoidal signals, limiting their ability to generate complex waveforms for challenging terrains. Neural networks and reinforcement learning have been incorporated to dynamically modulate CPG outputs. In one study, a neural network combined with tilt-angle feedback learned to generate target rhythmic signals from Matsuoka CPG outputs, enabling adaptive walking on slopes for a humanoid bionic robot. Reinforcement learning has also been used to tune CPG parameters like amplitude and frequency based on state feedback from soft snake-like bionic robots, where CPG directly controls pneumatic actuators. The state variables include distance to target, velocity, angular deviation, and joint curvature, as expressed in:
$$
s = [d, v, \theta, \dot{\theta}, \kappa_1, \kappa_2, \dots, \kappa_n]
$$
where \( d \) is distance, \( v \) is velocity, \( \theta \) is angle deviation, \( \dot{\theta} \) is its rate of change, and \( \kappa_i \) are joint curvatures. The reinforcement learning module outputs tuning inputs and frequency ratios to adjust CPG signals, facilitating robust locomotion in dynamic environments. These integrations expand CPG’s applicability for complex motion generation in bionic robots.
Conclusion
In this review, I have explored the advancements in CPG-based control for bionic robots, covering mathematical models, applications, and technological integrations. CPG’s bio-inspired approach offers significant benefits for bionic robot locomotion, including real-time adaptability, simplicity, and model-free operation. The evolution from joint-space to workspace trajectory generation has streamlined CPG networks, while its role has expanded from planner to direct controller for actuators like pneumatic muscles, aligning closely with biological neuromuscular coupling. Multi-objective optimization and intelligent algorithms have further enhanced CPG’s capability to generate diverse and complex motion signals, addressing limitations in parameter tuning and waveform modulation. Looking ahead, key research directions include improving multi-modal flexibility, dynamic environment adaptation, and energy efficiency in bionic robots. As CPG control continues to fuse with cutting-edge technologies, its potential to drive innovation in bionic robot design and application remains vast, promising more lifelike and resilient robotic systems.