This study introduces an integrated motion-control framework for achieving robust locomotion on low-cost quadruped robots. The framework combines biologically inspired Central Pattern Generators (CPGs) as structured gait priors, a leg-level multi-agent residual coordination policy, and hardware-aware adaptations including actuator abstraction and safety-aware command projection. Key low-cost design choices include a single shared actor with 39,560 parameters (≈155 KiB in 32-bit precision) evaluated at 50 Hz using compact proprioceptive observations, and avoidance of high-rate centralized optimizers, external motion capture, or heavy perception stacks on the physical robot. The authors validate the approach in simulation and on a physical low-cost quadruped across obstacles, ramps, stairs, and uneven terrain, reporting improvements in locomotion stability, energy efficiency, and terrain adaptability compared with baseline controllers.
Quadruped robots offer mobility advantages on uneven and unstructured terrain, making them relevant to inspection, search and rescue, and field exploration. Although high-performance quadrupeds exploit premium actuators, accurate sensing, and substantial computation to achieve dynamic behaviors, many open and affordable platforms face limitations in actuator bandwidth, sensor quality, and onboard processing. These constraints can degrade gait stability, disturbance rejection, and terrain adaptability.
Central Pattern Generators (CPGs) provide a compact, biologically inspired mechanism to generate rhythmic limb motion with low computational cost and high interpretability, which suits hardware-constrained platforms. However, pure CPG-based approaches may struggle with strongly varying terrain and the uncertainties of inexpensive sensors and actuators. Model-based and learning-based methods deliver advanced dynamic performance but often assume richer sensing, computation, or extensive simulation-to-real transfer procedures that low-cost systems lack.
The authors identify a gap between three research lines: rhythmic CPG control, distributed or modular coordination, and practical low-cost hardware adaptation. They propose a CPG-conditioned leg-level residual coordination framework to bridge these lines, preserving modularity and improving tolerance to per-leg variation while maintaining deployability on constrained hardware.
The proposed controller couples a CPG oscillator with a multi-agent residual policy using a shared-parameter actor architecture. The CPG provides explicit rhythm-related signals—phase, amplitude, contact schedule, step frequency, swing height, and nominal joint trajectories—which serve as structured priors for leg-level agents. Each leg agent receives the CPG state and local proprioceptive observations and outputs bounded residual corrections to the nominal commands.
A centralized training paradigm with a shared critic is employed during learning, but the runtime deployment on robot hardware uses only the lightweight shared actor evaluated at 50 Hz. The architecture aims to limit onboard memory and computation while preserving coordination across legs through shared parameters and a global state bus.
The central idea is to make the CPG oscillator state an explicit coordination prior for the residual multi-agent policy rather than treating CPGs and learned policies as independent modules. By exposing oscillator variables to each leg agent, the system encodes gait timing and swing/stance scheduling directly into the agent observations. Each leg then learns residual actions that correct nominal CPG-generated trajectories to account for local disturbances, actuation nonidealities, and terrain interactions.
Modeling each leg as an agent supports modularity and improved fault tolerance: the shared policy weights impose coherence while allowing legs to react to local state differences. The bounded residual approach constrains learned corrections to remain close to interpretable, safe nominal trajectories supplied by the CPG.
Design choices emphasize deployability on inexpensive platforms. The low-cost merit is highlighted by the single shared actor with 39,560 parameters (approximately 155 KiB in 32-bit precision) running at 50 Hz from compact proprioceptive inputs. The runtime footprint intentionally omits components that are often unavailable on low-cost robots: centralized high-frequency optimization, onboard external motion capture, sophisticated vision-based perception, direct torque sensing at high rates, and heavy online dynamics optimization are not required on the physical robot for deployment.
Actuator abstraction layers and safety-aware command projection are included to handle limited actuator bandwidth, backlash, and noisy sensing encountered in low-cost hardware. These abstractions map residual outputs to feasible motor commands while enforcing safety bounds.
The framework was validated in both simulation and on a physical low-cost quadruped. Test scenarios included traversal of obstacles, ramps, stairs, and uneven terrain to assess robustness across common locomotion challenges. The experiments compared the integrated CPG-conditioned multi-agent controller with baseline controllers to evaluate improvements in stability, energy use, and terrain adaptability.
The article indicates that the validation covered both simulated environments and real-world deployment on a low-cost robot, demonstrating the practicality of the approach for hardware-constrained platforms.
Experimental results reported by the authors show that combining a structured gait prior (CPG), a lightweight residual multi-agent coordination layer, and hardware-aware deployment enhances locomotion performance on low-cost quadrupeds. Specifically, the integrated system improved locomotion stability, energy efficiency, and terrain adaptability relative to baseline controllers in the tested scenarios. The authors present figures and tables that document quantitative and qualitative comparisons; details are available in the original article and its figures/tables.
The study frames the proposed controller as a bridge between biologically inspired rhythmic generation, decentralized coordination, and practical low-cost hardware adaptation. The authors argue that conditioning residual coordination on explicit CPG oscillator state preserves interpretability and provides robust priors, while the multi-agent residual policy compensates for real-world uncertainties and hardware limitations.
The paper also situates the work within broader threads of legged robotics research: CPGs for low-cost controllers, learning/optimization for dynamic behavior, and state-estimation challenges on inexpensive platforms. The authors note that pure CPG methods can be insufficient on strongly varying terrain and that learning-based residuals can help without imposing heavy runtime requirements.
Because the source article contains more exhaustive results, implementation specifics, and figures, readers should consult the original paper for algorithmic parameters, training details, and per-experiment metrics.
The article indicates that the minimal data set and accompanying code are available on GitHub at the repository provided in the source. Funding support was acknowledged from the Collaborative Innovation Center for Electrical Engineering and Intelligent Automation, with a grant noted for the lead author. The authors declared no competing interests.
The source article includes extensive references covering CPGs, model-based and learning-based locomotion, low-cost quadruped platforms, and state-estimation techniques relevant to legged robots. For exact citations and to review the referenced literature, consult the original PLoS ONE article.