Control methods for optimizing the power consumption of quadruped robot motors by utilizing the resonance characteristics of springs
By using multi-source perception hierarchical coding and multi-network collaborative reasoning, the quadruped robot estimates the instantaneous natural frequency of the parallel springs in the knee joint online and generates motor action commands, solving the problem of high motor power consumption in existing technologies and achieving stable energy-saving effects and reduced motor energy consumption.
Patent Information
- Application Number
- CN202611190695.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-06
- Publication Date
- 2026-11-10
AI Technical Summary
Existing quadruped robots, in their parallel elastic spring systems at the knee joints, struggle to adjust their step frequency in real time to match the resonance range, resulting in high motor power consumption and insufficient battery life. Current adaptive control and reinforcement learning solutions fail to effectively integrate body state and terrain information, making it impossible to stably achieve resonance-based power consumption reduction.
By combining multi-source perception hierarchical coding with multi-network collaborative reasoning, the robot acquires sensory data on its body, environment, and foot force. The instantaneous natural frequency of the parallel elastic element of the knee joint is estimated online using a pre-trained state evaluation network and a spring natural frequency evaluation network, generating motor action commands that enable the quadruped robot to track the current instantaneous natural frequency and reduce motor power consumption.
It achieves stable energy-saving performance for quadruped robots under different terrains and movement speeds, reduces motor energy consumption, is compatible with high-frequency control of onboard hardware, eliminates the need for additional frequency detection hardware, and is compatible with complex terrains and various movement conditions.
Smart Images

Figure CN123074380A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the technical field of quadruped robot control, and in particular to a control method and system for optimizing the power consumption of quadruped robot motors by utilizing the resonant characteristics of springs. Background Technology
[0002] Quadruped robots can adapt to various unstructured and complex terrains, exhibiting outstanding operational mobility. However, the high energy consumption of their drive motors and insufficient range limit their large-scale application. Existing energy-saving solutions use parallel elastic springs at the knee joints to form elastic actuators, relying on the resonance of the spring-mass system to share the load, thereby reducing the peak torque output of the motor. However, the leg elastic system is a nonlinear time-varying system; real-time changes in leg posture, ground stiffness, and leg load continuously alter the system's inherent resonant frequency. Existing adaptive control and model predictive control rely on precise dynamic models, which are complex and cannot accurately estimate the dynamically changing instantaneous inherent frequency online, nor can they continuously adjust the gait frequency to match the resonance range, thus limiting energy-saving effects. Reinforcement learning offers the advantage of model-free adaptive control, enabling it to autonomously learn optimal gait strategies. However, existing solutions lack a multi-source perception fusion mechanism and do not simultaneously combine body state, terrain information, and foot-end interaction mechanics to predict the instantaneous inherent frequency of the springs in real time, thus failing to stably achieve resonance-based energy saving.
[0003] Therefore, there is an urgent need for a control method and system that utilizes the resonant characteristics of springs to optimize the power consumption of quadruped robot motors. Summary of the Invention
[0004] To address the aforementioned technical problems, this application provides a control method and system for optimizing the power consumption of a quadruped robot motor by utilizing the resonance characteristics of a spring.
[0005] A first aspect of this application provides a control method for optimizing the power consumption of a quadruped robot motor by utilizing the resonance characteristics of a spring, comprising:
[0006] S101: Acquire the perception data of the quadruped robot, including body perception data, external environment perception data, and foot force perception data;
[0007] S102: Input the external environment perception data into the external perception encoder to obtain the implicit environmental representation vector;
[0008] S103: Input the ontology perception data and the environment implicit representation vector into the pre-trained state evaluation network to obtain the state implicit representation vector of the robot's overall characteristics;
[0009] S104: Input the foot force sensing data, the body sensing data and the environment implicit expression vector into the pre-trained spring natural frequency evaluation network, estimate and output the current instantaneous natural frequency of the passive elastic element at the joint online;
[0010] S105: Input the current instantaneous inherent frequency, the state implicit expression vector, and the ontology perception data into the pre-trained main network to generate motor action commands;
[0011] S106: Control the operation of the quadruped robot's motor according to the motor action command, so that the quadruped robot's step frequency tracks the current instantaneous natural frequency, thereby reducing motor power consumption;
[0012] The quadruped robot has elastic springs connected in parallel at the knee joint of each leg.
[0013] A second aspect of this application provides a control system for optimizing the power consumption of a quadruped robot motor by utilizing the characteristics of spring resonance, comprising:
[0014] The perception data module is used to acquire the perception data of the quadruped robot, including body perception data, external environment perception data, and foot force perception data.
[0015] The environmental terrain module is used to input the external environmental perception data into the external perception encoder to obtain the environmental implicit representation vector;
[0016] The state evaluation module is used to input the ontology perception data and the environment implicit expression vector into a pre-trained state evaluation network to obtain the state implicit expression vector of the robot's overall characteristics.
[0017] The natural frequency module is used to input the foot force sensing data, the body sensing data and the environmental implicit expression vector into a pre-trained spring natural frequency evaluation network, and to estimate and output the current instantaneous natural frequency of the passive elastic element at the joint online.
[0018] The motor instruction module is used to input the current instantaneous natural frequency, the implicit state representation vector, and the ontology perception data into the pre-trained main network to generate motor action instructions;
[0019] The tracking control module is used to control the operation of the motor of the quadruped robot according to the motor action command, so that the step frequency of the quadruped robot tracks the current instantaneous natural frequency and reduces the power consumption of the motor.
[0020] The quadruped robot has elastic springs connected in parallel at the knee joint of each leg.
[0021] A third aspect of this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the steps of the control method described above for optimizing the power consumption of a quadruped robot motor using spring resonance characteristics.
[0022] In a fourth aspect of this application, a computer-readable storage medium is provided, the computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described control method for optimizing the power consumption of a quadruped robot motor using the characteristics of spring resonance.
[0023] The beneficial effects of the control method and system for optimizing the power consumption of a quadruped robot motor by utilizing the resonance characteristics of springs provided in this application are as follows: This application estimates the instantaneous natural frequency of the parallel spring in the knee joint online by combining multi-source sensing hierarchical coding with multi-network collaborative reasoning, driving the gait frequency to follow the spring resonance frequency. Features are extracted hierarchically from three types of sensing data: body, environment, and foot force, respectively constructing latent vectors for the environment and robot state, accurately coupling the real-time dynamic characteristics of the elastic element. Unlike offline fixed-frequency lookup table methods, this method can dynamically update the resonance reference according to terrain and body load, using spring resonance to offset most of the joint load torque, directly reducing the motor output torque. The entire control process operates in a closed loop in real time, adapting to high-frequency control of onboard hardware without the need for additional frequency detection hardware, relying solely on existing sensors to complete the entire process calculation, reducing motor energy consumption during continuous quadruped robot movement, and being compatible with various terrains and movement speeds, with stable and reliable energy-saving effects. Attached Figure Description
[0024] Figure 1 A flowchart illustrating a control method for optimizing the power consumption of a quadruped robot motor using the resonant characteristics of a spring, provided in an embodiment of this application;
[0025] Figure 2 A diagram of the Actor-Critic training framework for a control system that optimizes the power consumption of a quadruped robot motor using the characteristics of spring resonance, provided in an embodiment of this application.
[0026] Figure 3 A diagram illustrating the strategy network (Actor) of a control system that optimizes the power consumption of a quadruped robot motor using the characteristics of spring resonance, provided in an embodiment of this application.
[0027] Figure 4 A structural block diagram of a control system for optimizing the power consumption of a quadruped robot motor by utilizing the resonant characteristics of a spring, provided in an embodiment of this application;
[0028] Figure 5 This is a hardware architecture diagram of a control system for optimizing the power consumption of a quadruped robot motor by utilizing the resonant characteristics of a spring, provided in an embodiment of this application.
[0029] Figure 6 This is a schematic block diagram of an electronic device provided in an embodiment of this application. Detailed Implementation
[0030] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.
[0031] To make the purpose, technical solution, and advantages of this application clearer, the following will be described in conjunction with the appendix. Figure 1 - Appendix Figure 6 The following is an explanation using specific examples.
[0032] Please refer to Figure 1 , Figure 1 This is a flowchart illustrating a control method for optimizing the power consumption of a quadruped robot motor using the resonant characteristics of a spring, according to an embodiment of this application. The method includes:
[0033] S101: Acquire the perception data of the quadruped robot, including body perception data, external environment perception data, and foot force perception data.
[0034] In this embodiment, the body perception data is the intrinsic sensing data collected from the robot's body and joints to characterize the robot's posture, joint motion, and motor operating status, representing the robot's own dynamic characteristics. This includes: IMU (Inertial Measurement Unit): body posture quaternion, pitch / roll / yaw angle, three-axis body linear velocity, three-axis angular velocity, and body acceleration; joint encoder: real-time angles and angular velocities of the hip and knee joints of the four legs, including multi-step historical time-series data; motor information: drive current and output torque of each joint; and auxiliary status: motion phase and body height above the ground.
[0035] In this embodiment, the external environment perception data is the raw data of the robot's external terrain environment collected by vision / laser sensors. It is used to characterize the ground topography, roughness, stiffness, slope, and obstacles, providing environmental mechanics priors for spring frequency estimation. This includes: depth maps from depth cameras, dense terrain point clouds, terrain height sampling data, ground texture, slope / rugged terrain features, and obstacle location information.
[0036] In this embodiment, the foot force sensing data is the mechanical data of the foot-ground contact interaction collected by a six-dimensional foot force sensor. It is the core input for determining ground stiffness, damping, real-time leg load, and online estimation of the instantaneous natural frequency of the spring. It includes: three-dimensional foot contact force Fx / Fy / Fz, three-dimensional torque Mx / My / Mz, foot contact state, plantar normal pressure, and leg impact load.
[0037] S102: Input the external environment perception data into the external perception encoder to obtain the implicit representation vector of the environment.
[0038] In this embodiment, the external perception encoder is a feature extraction sub-network with a convolutional neural network (CNN) at its core, serving as a pre-module of the Actor strategy network. It takes raw external environment perception data as input, performs spatial feature extraction, information dimensionality reduction, and redundancy removal, and outputs latent vectors of a unified dimension. It can automatically extract latent mechanical features such as terrain roughness, ground stiffness, ground damping, slope, and terrain type, eliminating the need for manual feature design. It also unifies the dimension of output data from different sensors, adapting to the input format of subsequent networks.
[0039] In this embodiment, the environmental implicit representation vector is a low-dimensional dense numerical vector output by the external perceptual encoder. It is an abstract representation of the terrain environment features and has no intuitive physical image meaning. However, the numerical values of each dimension of the vector implicitly contain environmental mechanical properties such as ground stiffness, roughness, damping, and slope. It can standardize intermediate features and serve as a shared input for the state evaluation network and the spring natural frequency evaluation network, providing environmental constraints for online estimation of the instantaneous natural frequency of the spring.
[0040] S103: Input the ontology perception data and the implicit representation vector of the environment into the pre-trained state evaluation network to obtain the implicit representation vector of the robot's overall characteristics.
[0041] In this embodiment, the pre-trained state evaluation network is a sub-network within the Actor policy network, composed of multi-layer fully connected neural networks, and is pre-trained offline on a multi-terrain simulation dataset. It inputs ontology sensing sequences and environmental latent vectors to achieve cross-modal feature fusion. This enables the integration of two types of heterogeneous data: the robot's own state and external terrain data. It automatically mines the overall dynamic features of the robot under the coupling of body posture, leg load, and ground conditions, eliminating the dimensional mismatch problem between the two types of data and outputting a unified global state representation.
[0042] In this embodiment, the implicit state representation vector (IVV) of the robot's overall characteristics is a fixed-dimensional dense feature vector output by the state evaluation network. It is an abstract representation of the robot's overall dynamic state. It includes global characteristics such as body movement speed, body stability, real-time leg load distribution, terrain and body coupling dynamics, and system energy storage status. Instead of separately separating individual joint / terrain information, the fused global features are provided to the main network to generate matching resonance motor commands.
[0043] S104: Input foot force perception data, proprioceptive perception data and environmental implicit representation vectors into a pre-trained spring natural frequency evaluation network to estimate and output the current instantaneous natural frequency of the passive elastic element at the joint online.
[0044] In this embodiment, the pre-trained spring natural frequency evaluation network is a temporal recurrent neural network (LSTM / GRU) built into the Actor policy network, which is pre-trained offline under supervision on multi-terrain and multi-load simulation datasets. During the training phase, the theoretical equivalent inertia calculated by the parallel axis theorem and the natural frequency obtained by solving the spring stiffness are used as labels to supervise the network. It can model multi-source temporal sensor data, capture the dynamic time-varying characteristics caused by leg landing, airborne, and attitude changes, correct ground interactions and load disturbances that cannot be fitted by pure theoretical models in real time, and output accurate real-time resonant frequencies.
[0045] In this embodiment, the passive elastic element at the joint refers to the torsion spring installed in parallel at each knee joint. It is a passive energy storage structure that realizes the mutual conversion of kinetic energy and elastic potential energy through deformation, and is the hardware basis for resonant energy saving.
[0046] In this embodiment, the current instantaneous natural frequency is the real-time equivalent natural oscillation frequency of the knee joint spring-linkage equivalent mass system under the combined effects of the current body posture, leg load, and ground environment. This value continuously and dynamically changes with the motion state and is not a fixed constant. The controller adjusts the robot's walking step frequency with this frequency as the target, so that the system enters the resonance range, reducing the motor output torque and power consumption.
[0047] S105: Input the current instantaneous natural frequency, implicit state representation vector, and ontology perception data into the pre-trained main network to generate motor action commands.
[0048] In this embodiment, the pre-trained main network is the final decision sub-network of the Actor policy network, which is composed of multi-layer fully connected neural networks. It is trained uniformly in a multi-terrain simulation environment through the Actor-Critic reinforcement learning framework. It is the core decision module for generating motor control quantities. It can integrate multi-dimensional information such as resonance frequency, overall machine state, and real-time ontology observation, while taking into account multiple objectives such as speed tracking, fuselage stability, resonance matching, and torque constraints, and outputs continuous motor motion control quantities.
[0049] In this embodiment, the motor motion commands are the final control outputs from the main network, which are sent to the motor drivers of each leg joint for execution. These commands include: target joint angle, target joint torque, joint rotation speed, etc.; they directly determine the robot's step frequency, stride length, and body posture, and are used to drive the robot's gait to match the instantaneous natural frequency of the springs.
[0050] S106: Controls the operation of the quadruped robot's motors according to the motor action commands, so that the quadruped robot's step frequency tracks the current instantaneous natural frequency, reducing motor power consumption.
[0051] In this embodiment, the motor action command is a control output quantity output by the main network. It can be a quantified control signal such as the target torque, target joint angle, and target speed of each joint, which is directly sent to the motor driver and serves as the basis for the motor to execute motion.
[0052] In this embodiment, controlling the motor operation of the quadruped robot means that the robot's onboard controller sends motor action commands to the corresponding motor drivers of the hip and knee joints of the four legs. The drivers adjust the output torque / rotation angle of the motors according to the commands, thereby driving the leg links to complete the entire gait movement of swinging, supporting, and taking off.
[0053] In this embodiment, step frequency is a core parameter of the quadruped robot's gait, representing the number of times a single leg completes a landing-air cycle per unit time, measured in Hz; step frequency directly determines the oscillation rhythm of the leg's back-and-forth swing.
[0054] In this embodiment, tracking the current instantaneous natural frequency is achieved by using the real-time estimated instantaneous natural frequency of the spring as the target reference, adjusting the gait cycle and gait speed, so that the oscillation frequency of the robot's leg reciprocating motion is consistent with the natural oscillation frequency of the spring-linkage system, thereby enabling the system to enter a resonant state.
[0055] In this embodiment, the current instantaneous natural frequency is the real-time equivalent oscillation frequency of the elastic system composed of the knee joint parallel spring and the leg linkage under the current terrain, leg load, and fuselage attitude, which changes dynamically with the motion state.
[0056] In this embodiment, the reduction of motor power consumption occurs when the spring stores elastic potential energy during the support phase and releases potential energy during the swing phase under resonant conditions, thus offsetting most of the inertial load and impact load of the legs. The motor does not need to continuously output large peak torque, and the average output power and average current of the motor decrease, thereby reducing the overall energy consumption of the machine.
[0057] See Figure 3 , Figure 3 The illustration for the strategy network (Actor) provided in one embodiment of this application includes: inputting external environment perception data into an external perception encoder to obtain an environmental implicit expression vector; inputting body perception data and the environmental implicit expression vector into a state evaluation network to obtain a state implicit expression vector; inputting foot force perception data, body perception data, and the environmental implicit expression vector into a spring natural frequency evaluation network to estimate and output the current instantaneous natural frequency of the elastic spring online; inputting the current instantaneous natural frequency, the state implicit expression vector, and the body perception data into a main network to generate motor action commands and send them to the motor drive system, so that the quadruped robot's movement step frequency tracks the current instantaneous natural frequency and reduces motor power consumption by utilizing spring resonance.
[0058] As can be seen from the above, this application estimates the instantaneous natural frequency of the parallel spring in the knee joint online by combining multi-source perception hierarchical coding with multi-network collaborative reasoning, driving the gait frequency to follow the spring's resonant frequency. Features are extracted hierarchically from three types of perception data: body, environment, and foot force, respectively constructing latent vectors for the environment and robot state, accurately coupling the real-time dynamic characteristics of the elastic element. Unlike offline fixed-frequency lookup methods, it can dynamically update the resonant reference according to terrain and body load, utilizing spring resonance to offset most of the joint load torque, directly reducing the motor output torque. The entire control process operates in a closed-loop real-time manner, adapting to high-frequency control of onboard hardware without requiring additional frequency detection hardware, relying solely on existing sensors to complete the entire process calculation, reducing motor energy consumption during continuous quadruped robot movement, while also being compatible with various terrains and movement speeds, with stable and reliable energy-saving effects.
[0059] In one embodiment of this application, external environment perception data is input into an external perception encoder to obtain an implicit environmental representation vector, including:
[0060] The external perception encoder uses a convolutional neural network to extract features and compress dimensions from external environment perception data.
[0061] The environmental implicit representation vector is used to characterize one or more of the following: terrain type, terrain roughness, ground stiffness characteristics, and ground damping characteristics.
[0062] In this embodiment, a convolutional neural network (CNN) is a deep learning network with the ability to extract local features. It extracts spatial local features such as terrain undulation, texture, and slope in images / point clouds by sliding convolutional kernels, and compresses data dimensions with pooling layers to adapt to visual environmental perception data.
[0063] In this embodiment, feature extraction automatically mines key information related to terrain mechanics from the original environmental data using convolutional kernels: ground undulations, rough textures, slope changes, soft and hard boundaries, etc., while removing irrelevant background interference information. Dimensional compression transforms the original image / point cloud data, which has tens of thousands of dimensions, into low-dimensional vectors of tens to hundreds of dimensions after mapping through convolution, pooling, and fully connected layers, reducing computational load and lowering onboard computing power consumption.
[0064] In this embodiment, the implicit environmental representation vector is a low-dimensional floating-point feature vector output by the encoder. It has no intuitive visual meaning, and the numerical values of each dimension of the vector implicitly quantify the terrain's mechanical properties. Terrain type distinguishes different ground scenarios such as flat hard ground, grass, sand, gravel, slopes, and rugged roads. Terrain roughness refers to the degree of unevenness of the ground surface; higher roughness results in greater impact upon foot landing, altering the system's equivalent damping. Ground stiffness characteristics describe the ease with which the ground deforms under pressure; greater stiffness leads to stronger equivalent foot support elasticity, directly changing the spring system's equivalent natural frequency. Ground damping characteristics represent the ground's ability to absorb kinetic impact energy; damping consumes spring oscillation energy, affecting resonance maintenance.
[0065] As can be seen from the above, this embodiment uses a convolutional neural network to compress and extract environmental perception data features through an external perception encoder. The generated implicit environmental representation vector uniformly represents multi-dimensional terrain attributes such as terrain roughness, ground stiffness, and damping, solving the problems of high dimensionality, redundant information, and inability to directly participate in network inference in the original environmental image / point cloud data. The unified structured implicit vector output can standardize the input to subsequent state evaluation and spring frequency evaluation networks, eliminating the model adaptation problem caused by differences in different terrain sensing data formats. The CNN's local feature extraction capability can accurately capture subtle mechanical features of the ground, providing accurate terrain mechanical priors for estimating the instantaneous natural frequency of the spring. This makes the frequency estimation results fit the real ground elastic constraints, improves the step frequency resonance matching accuracy, further enhances the motor power consumption reduction effect, and improves the robustness of the control algorithm under complex terrain.
[0066] In one embodiment of this application, the pre-trained spring natural frequency evaluation network is a recurrent neural network, comprising:
[0067] Recurrent neural networks, including LSTM or GRU units, model the temporal dynamic characteristics of input data to track and output the current instantaneous natural frequency online.
[0068] During the training phase, the equivalent rotational inertia of the joints is calculated in advance based on the mass, center of mass, inertia, motor rotor inertia, and reduction ratio of the robot links using the parallel axis theorem. The theoretical natural frequency is obtained through the spring stiffness coefficient and used as the training benchmark label for the spring natural frequency evaluation network.
[0069] In this embodiment, the pre-trained spring natural frequency evaluation network is a dedicated sub-network within the Actor policy network. It takes foot force, body, and environmental latent vectors as input and outputs the instantaneous natural frequency of the knee joint spring. Offline supervision and reinforcement learning are jointly pre-trained in a simulation environment, and after deployment, it performs direct real-time inference. Recurrent Neural Networks (RNNs) are neural networks adept at processing continuous temporal data, capable of remembering gait, force, and posture information from the previous moment, and are adapted to dynamic prediction of continuous robot motion.
[0070] In this embodiment, the LSTM / GRU units are two mainstream improved units of recurrent neural networks. By using a gating structure to preserve long-term temporal information, they solve the gradient vanishing problem of ordinary RNNs and can accurately capture the dynamic frequency shift caused by continuous changes in landing, airborne, and posture. The temporal dynamic characteristics of the input data are that sensor data such as foot force, joint angle, and body posture change continuously with the gait cycle. The leg load and equivalent inertia are different at different times, and the natural frequency fluctuates continuously. This temporal variation pattern is the temporal dynamic characteristic.
[0071] In this embodiment, the online tracking output of the current instantaneous natural frequency is the real-time inference of each control cycle during the robot's operation, and the resonant frequency is continuously updated according to changes in terrain, load, and attitude, rather than using a fixed offline theoretical value.
[0072] In this embodiment, the mass, center of mass, and moment of inertia of the link are inherent mechanical parameters of the robot's thigh, calf, and other leg components; the moment of inertia characterizes the rotational inertia of the link, and the position of the center of mass is used to calculate the moment of inertia using the parallel axis theorem.
[0073] In this embodiment, the motor rotor inertia and reduction ratio are the rotational inertia of the motor output shaft itself; the reduction ratio is the transmission ratio of the reducer, and the inertia of the motor side needs to be converted to the joint rotation axis.
[0074] In this embodiment, the parallel axis theorem is a fundamental formula in rigid body mechanics, used to uniformly convert the moment of inertia of the link that deviates from the joint rotation axis to the center of rotation of the knee joint, and then superimpose them to obtain the total equivalent moment of inertia.
[0075] In this embodiment, the equivalent rotational inertia of the joint is denoted as i. eq The moment of inertia, calculated from all leg links and motor rotors read from the simulation environment and superimposed to be equivalent to the knee joint axis, is the core parameter for calculating the natural frequency of the spring system. The spring stiffness coefficient, k, is the inherent parameter of the parallel torsion springs in the knee joint, characterizing the spring's resistance to deformation. The theoretical natural frequency, denoted as W, is the calculated standard resonant frequency, used as the ground truth for training the network. The formula for calculating the theoretical natural frequency is: W = .
[0076] In this embodiment, the training baseline label is the true value data that serves as the standard answer in deep learning supervised learning. During network training, the predicted frequency is approximated to the theoretical frequency, and physical prior constraints are introduced to avoid the prediction results from violating the laws of dynamics.
[0077] As can be seen from the above, this embodiment uses an LSTM / GRU recurrent unit to model temporal dynamics through a spring natural frequency evaluation network. This allows for continuous tracking of instantaneous natural frequency changes during motion, adapting to temporal conditions such as robot takeoff, landing, and sudden load changes. During the training phase, the theoretical natural frequencies calculated using the equivalent rotational inertia of links, motors, and reduction mechanisms serve as supervisory labels, providing the network with a precise physical benchmark and avoiding physically meaningless prediction biases that can occur in purely data-driven models. By integrating dynamic priors with temporal deep learning, and balancing physical interpretability with real-time online prediction capabilities, this embodiment can compensate for the equivalent inertia shift caused by ground forces in real time compared to purely offline dynamic calculations. It outputs accurate instantaneous resonant frequencies, providing reliable input for the main network to generate resonant matching gait, and maximizing the energy-saving effect of the spring.
[0078] In one embodiment of this application, a control method for optimizing the power consumption of a quadruped robot motor using the resonant characteristics of a spring further includes:
[0079] The external perception encoder, the state evaluation network, the spring natural frequency evaluation network, and the main network constitute the Actor policy network, which is trained through the reinforcement learning Actor-Critic framework.
[0080] During training, the input to the Critic network includes sensory data and privileged information, which includes one or more of the following: ground friction coefficient, ground stiffness parameters, leg collision status, and flight duration.
[0081] After training, only the Actor policy network is exported for deployment; the Critic network does not participate in inference computation during the deployment phase.
[0082] In this embodiment, the Actor policy network is a complete policy inference backbone network consisting of four types of sub-networks: external sensing encoder, state evaluation network, spring natural frequency evaluation network, and main network. Its function is to receive sensor observations and directly output motor control actions. It is the only network that runs during the deployment phase.
[0083] In this embodiment, the Actor-Critic framework is a classic reinforcement learning dual-network training architecture, which is divided into two parts: Actor (policy network) and Critic (value evaluation network) for collaborative training. The Actor is responsible for generating actions, and the Critic evaluates the quality of the current state. The network parameters are updated through the cooperation of the two.
[0084] In this embodiment, the Critic network is a value evaluation network used only during offline training and does not participate in real-machine operation. It takes observation and privileged information as input and outputs the current state value estimate, which is used to calculate gradient loss and optimize the Actor network. Perception data is a general term for three types of collected data: ontology perception data, external environment perception data, and foot force perception data, which are the common basic inputs for training and inference. Privileged information consists of low-level dynamic parameters that can only be directly read in the simulation environment and are not obtained in real time by the real robot without sensors. It only assists in training to improve model accuracy and cannot be obtained in real-machine deployment. Privileged information includes one or more of the following: ground friction coefficient, ground stiffness parameter, leg collision state, and flight duration. Among them, the ground friction coefficient is the magnitude of friction at the contact interface between the foot and the ground, which affects leg slippage and gait stability; the ground stiffness parameter is the ground elastic deformation coefficient, which directly changes the equivalent natural frequency of the spring system; the leg collision state is the Boolean / intensity signal of the impact between the leg link and the ground and the fuselage; and the flight duration is the duration of a single leg leaving the ground, representing the gait cycle and load distribution.
[0085] In this embodiment, the deployment phase is the stage after the network training is completed, where the model is ported to the onboard controller of a real quadruped robot for real-time closed-loop control of the robot's movement. Inference computation is the calculation process in which the network receives input data, performs forward computation, and outputs results; the Critic network only participates in inference during training and is not loaded or computed during actual machine operation.
[0086] See Figure 2 , Figure 2 The diagram below illustrates an Actor-Critic training framework provided in an embodiment of this application. The training framework includes an environment module (Env), an Actor policy network, a Critic value network, a GAE advantage estimation module, an updater, and a loss calculation and backpropagation module.
[0087] During training, the Actor receives observation data (obs) output by the Env and outputs action instructions (actions) to the Env. The Env updates its state based on the actions and returns a reward and the next observation. The Critic receives critical obs, which includes obs and privilege information, and outputs a value estimate of the current state (Value). The GAE module calculates the advantage estimate based on the value and reward sequence. The updater calculates the loss value (Loss) based on the advantage estimate, calculates the gradient through backward propagation, and performs policy updates to optimize the network parameters of the Actor and Critic. The updated Actor continues to interact with the environment, forming a training loop until the network converges.
[0088] After training, only the Actor policy network is exported for the real-time control deployment of the physical robot; the Critic network does not participate in the inference during the deployment phase.
[0089] As can be seen from the above, this embodiment integrates the encoder, evaluation network, and main network into an Actor policy network, and completes unified training based on the Actor-Critic reinforcement learning framework. During the training phase, privileged information such as ground friction and leg collision is introduced to assist the Critic network in estimation, compensating for the lack of single-sensor observation information and narrowing the difference between the simulation and real robot dynamics domains. In the deployment phase, only the lightweight Actor network is retained to perform inference, and the Critic estimation network is eliminated, reducing the computing power overhead of the onboard computing unit and meeting the requirements of high-frequency real-time control. The entire network completes the end-to-end perception encoding, frequency estimation, and motion generation inference without the need for segmented and independent scheduling of multiple algorithm modules, resulting in lower control latency and ensuring that the step frequency follows the spring resonance frequency in real time, thereby improving the real-time performance of the overall machine control while reducing motor power consumption.
[0090] In one embodiment of this application, the training process of the reinforcement learning Actor-Critic framework includes:
[0091] S201: Construct a dynamic model of a quadruped robot in a simulation environment and set up various training terrains;
[0092] S202: Input the perception data into the Actor policy network, the Actor policy network outputs action commands and applies them to the dynamic model of the quadruped robot in the simulation environment, driving the dynamic model of the quadruped robot to perform actions;
[0093] S203: Input the perception data and privileged information into the Critic network and output the value estimate of the current state;
[0094] S204: Calculate the reward value based on the state changes after the quadruped robot performs an action according to its dynamic model, and use the GAE algorithm to calculate the advantage estimate;
[0095] S205: Calculate policy loss and value function loss based on advantage estimation, and update the network parameters of the Actor policy network and Critic network through backpropagation via the optimizer;
[0096] S206: Repeat steps S202 to S205 until the network converges.
[0097] In this embodiment, the Actor-Critic framework is a dual-network training architecture for reinforcement learning. The Actor generates control actions, and the Critic evaluates the state value. The two work together to complete the iterative optimization of network parameters.
[0098] In this embodiment, the simulation environment is a high-fidelity physical simulation platform, such as Isaac Lab or MuJoCo, which can simulate real rigid body dynamics, ground mechanics, and robot motion for offline training, avoiding repeated debugging and hardware wear and tear on the actual machine.
[0099] In this embodiment, the dynamic model of the quadruped robot is a digital URDF model, which fully reproduces the real mechanical dynamic characteristics of the robot, such as the mass of the robot links, inertia, motor reduction mechanism, stiffness of the parallel springs in the knee joint, and joint friction.
[0100] In this embodiment, various training terrains are simulated ground surfaces with different mechanical properties, including flat land, sand, grass, slopes, and gravel rugged roads, covering different ground stiffness, damping, and roughness to improve the model's generalization ability.
[0101] In this embodiment, the Actor policy network is an integrated inference network that combines an external sensing encoder, a state evaluation network, a spring natural frequency evaluation network, and a main network. It takes observations as input and outputs motor motion commands. These commands are target torque / angle control values for each joint, which are sent to the simulation dynamics model to drive leg movement. Acting on the simulation dynamics model means inputting the network's output motion into the simulation engine to update the robot's joint drive values. The executed motion is the robot's legs completing periodic gait movements of support, swing, and takeoff within the simulation.
[0102] In this embodiment, the Critic network is a dedicated value evaluation network used for training and is not deployed on the actual aircraft. It is input with sensor data and privileged information. Privileged information consists of low-level dynamic parameters obtainable only through simulation and collected by sensors on the actual aircraft, including ground friction coefficient, ground stiffness, leg collision status, and flight duration. The value estimate of the current state is a scalar value output by the Critic, representing the expected long-term cumulative reward under the current observation state, used to determine the quality of the state.
[0103] In this embodiment, the state change after the action is the change in body posture, joint angle, foot contact force, and terrain interaction conditions after the robot completes one gait step. The reward value is a quantitative feedback calculated in real time by a preset multi-objective composite reward function, with positive incentives for speed tracking and resonance matching, and negative penalties for large torque and body tilt.
[0104] In this embodiment, the GAE algorithm is a standard reinforcement learning algorithm that combines multi-step reward and value estimation to calculate the advantage function, thereby reducing policy update variance and improving training stability.
[0105] Specifically, the GAE algorithm data consists of five layers from input to output. The first layer collects the temporal reward generated after the simulation execution action and the output state value of the Critic network. The second layer calculates the single-step temporal difference residual frame by frame. The third layer introduces a discount factor to attenuate the long-term residual. The fourth layer obtains the comprehensive advantage by weighting and accumulating the multi-step residuals through the GAE attenuation factor. The fifth layer outputs the final advantage estimate to supply the loss calculation stage.
[0106] The inputs to the GAE algorithm include a continuous time-series immediate reward sequence, the time-series state value output by the Critic network, trajectory termination markers, and two types of preset hyperparameters: reward discount and GAE decay. The GAE output is a unique scalar advantage estimate, which quantifies the additional gain of the current action relative to the average action. This value is directly used to calculate the policy loss and value function loss, and to complete the reverse update of the parameters of the Actor and Critic networks.
[0107] In this embodiment, advantage estimation represents the additional reward that the current action can obtain compared to the average action, and it is the core basis for updating the Actor network.
[0108] In this embodiment, the policy loss is the Actor network loss, which measures the deviation between the policy output action and the optimal resonant energy-saving action, and is used to update the policy network.
[0109] In this embodiment, the value function loss is the Critic network loss, which measures the deviation between the Critic output value estimate and the true long-term reward. The optimizer is a gradient descent optimization tool, such as Adam, which automatically calculates gradients and updates network weights.
[0110] In this embodiment, backpropagation is a standard parameter update method in deep learning, where gradients are passed backward from the loss value layer by layer to correct the network weight parameters. The network parameters are the weights and biases of each layer of the Actor and Critic neural networks, which are continuously iterated and optimized during the training process.
[0111] In this embodiment, repeatedly executing S202-S205 represents continuously collecting multiple rounds of simulated interaction samples for batch iterative training. Network convergence occurs after multiple consecutive iterations when the fluctuation amplitudes of the policy loss and value function loss are less than the preset policy loss threshold and preset value function loss threshold, respectively, and the mean reward stabilizes and no longer increases significantly, indicating that policy learning is complete.
[0112] Specifically, the method for determining the preset policy loss threshold is as follows: conduct small-batch pre-training in the simulation environment, collect data on the changes in policy loss under multiple iterations, and after the reward growth slows down significantly and the model enters the quasi-convergence interval, calculate the maximum difference in the single-round fluctuation of policy loss in multiple consecutive training rounds, and add a 5%-10% fault tolerance margin on this basis to obtain the final preset policy loss threshold, which is used to determine whether the fluctuation of policy loss in the later stage of training has reached the stable standard.
[0113] The method for determining the preset value function loss threshold is as follows: First, perform pre-training iterations in the simulation environment. When the reward growth tends to level off and the model enters the quasi-convergence stage, statistically analyze the fluctuation difference of the intrinsic value function loss in each round of continuous training. Extract the maximum fluctuation amplitude of the loss in this interval, and then increase it by 5%-10% to obtain the preset value function loss threshold, which serves as the judgment criterion for the stability of the network output.
[0114] As can be seen from the above, this embodiment completes multi-terrain iterative training by building a complete quadruped dynamics model in a simulation environment, accurately calculates the state-action advantage through GAE generalized advantage estimation, optimizes the policy loss and value function loss respectively, and updates the Actor and Critic network parameters bidirectionally and synchronously, thereby improving the network convergence speed and training stability. The complete closed-loop simulation training process eliminates the need for repeated debugging on real devices, reducing hardware consumption and R&D costs; the multi-terrain samples cover various stiffness and damping conditions, and the trained network has extremely strong generalization ability, enabling it to quickly adapt to unknown terrains after deployment on real devices. Multiple rounds of iterative optimization until network convergence ensure stable policy output and resonant matching gait, continuously suppressing motor output torque, and guaranteeing long-term low-power motion performance from the source of algorithm training.
[0115] In one embodiment of this application, calculating a reward value based on the state change after performing an action according to the dynamic model of the quadruped robot includes:
[0116] The reward value is calculated by a reward function, which includes several of the following reward items: linear velocity tracking reward item, angular velocity tracking reward item, resonance recognition reward item, joint torque penalty item, and motion stability penalty item;
[0117] The linear velocity tracking reward is used to incentivize the tracking of the desired linear velocity in the horizontal direction based on the deviation between the robot's actual linear velocity and the desired linear velocity.
[0118] The angular velocity tracking reward is used to incentivize tracking of the desired yaw angular velocity based on the deviation between the robot's actual angular velocity and the desired angular velocity.
[0119] The resonance recognition reward is used to incentivize the movement cadence to track the current instantaneous natural frequency based on the deviation between the current instantaneous natural frequency and the current movement cadence.
[0120] The joint torque penalty term is used to constrain the motor output torque based on the amplitude of the output torque of each joint motor.
[0121] The motion stability penalty is used to penalize the fuselage tilt angle for exceeding the allowable range, based on the deviation of the fuselage roll and pitch angles.
[0122] In this embodiment, the state change after the dynamic model performs an action represents the updated overall state of the simulated robot after completing a round of gait actions. This includes real-time observation data such as body speed, yaw rate, movement step frequency, output torque of each joint, and body pitch / roll angle, which serve as the input basis for calculating rewards.
[0123] In this embodiment, the reward value is a scalar value output by the reward function in a single control cycle. A positive value represents excellent action performance, while a negative value represents a defective action. It is used for GAE advantage estimation, network loss calculation, and to drive the strategy towards energy-saving and stable gait optimization.
[0124] In this embodiment, the reward function is a quantitative calculation formula that integrates multiple types of rewards and penalties, uniformly quantifying the quality of the robot's motion performance, and serving as a mechanism for reinforcement learning to guide the network's learning objectives.
[0125] In this embodiment, the linear velocity tracking reward (positive incentive) is calculated based on the difference between the actual horizontal x / y linear velocity of the torso and the desired linear velocity issued by the user. The smaller the deviation, the higher the reward, prompting the robot to move forward and laterally with precision. The actual linear velocity is the current horizontal movement speed of the robot body calculated by the IMU; the desired linear velocity is the target movement speed command issued by the upper layer.
[0126] In this embodiment, the angular velocity tracking reward (positive incentive) is calculated based on the deviation between the actual yaw angular velocity and the desired turning angular velocity of the robot. The smaller the deviation, the higher the reward, ensuring that the robot accurately completes the turning maneuver. The actual angular velocity is the real-time rotational angular velocity of the robot's z-axis; the desired angular velocity is the target turning command input externally.
[0127] In this embodiment, the resonance recognition reward (positive core incentive) is calculated based on the difference between the real-time estimated instantaneous natural frequency of the spring and the robot's actual step frequency. The closer the step frequency is to the natural frequency, the higher the reward, guiding the controller to actively match the resonance range and reduce power consumption. The current instantaneous natural frequency is the real-time resonance frequency of the knee joint spring output by the network; the step frequency is the number of complete leg landing-swing cycles per unit time.
[0128] In this embodiment, the joint torque penalty term (negative constraint) is calculated based on the real-time output torque amplitude of all hip and knee joint motors. The larger the torque, the larger the penalty value, which suppresses the motor output of large torque and reduces energy consumption from the source.
[0129] In this embodiment, the motion stability penalty (negative constraint) is calculated based on the offset of the fuselage pitch and roll angles relative to the horizontal reference. The greater the degree of fuselage tilt, the higher the penalty, preventing the robot from becoming unbalanced and tipping over in pursuit of energy conservation through resonance. The roll angle is the left-right tilt angle of the fuselage; the pitch angle is the forward-backward pitch angle of the fuselage.
[0130] As can be seen from the above, this embodiment integrates multiple reward items such as speed tracking, resonance matching, torque constraint, and stability penalty through a reward function to form a multi-objective collaborative optimization mechanism. Speed-related rewards ensure that the robot accurately executes movement and turning commands, satisfying basic motion functions; resonance recognition rewards specifically guide the network to match the step frequency with the spring's natural frequency, reducing motor power consumption; torque and stability penalties constrain large torque output and significant body tilt, avoiding instability and overload problems caused by pursuing resonance energy saving. The multi-dimensional rewards mutually restrain each other, preventing the network from unilaterally pursuing a single indicator, and taking into account the three major requirements of motion command tracking, resonance energy saving, and body stability. This achieves smooth and controllable motion under the premise of low power consumption, avoiding the defects of gait distortion and motor overload caused by a single reward.
[0131] In one embodiment of this application, a control method for optimizing the power consumption of a quadruped robot motor using the characteristics of spring resonance is provided, wherein the reward function includes:
[0132] The reward function is R = α·(r_lv + r_av) + β·r_r + γ·r_τ + δ·r_stab;
[0133] Wherein, α, β, γ, and δ are the weight coefficients corresponding to each reward item, α > 0, β > 0, γ < 0, and δ < 0; r_lv is the linear velocity tracking reward, used to constrain the robot's torso's x and y direction travel speed to match the desired command; r_av is the yaw rate tracking reward, used to constrain the body's turning rate to match the desired command; r_r is the resonance recognition reward, used to guide the frequency estimate output by the network to match the real-time equivalent natural frequency of the joint; r_τ is the joint torque penalty term, used to suppress the motor's output of large torque to reduce power consumption; r_stab is the motion stability penalty term; the reward function is supplemented with a body tilt penalty term to ensure the robot's motion stability.
[0134] In this embodiment, the reward function R is the total scalar reward value output in a single control cycle, which comprehensively measures the quality of the robot's action in that step. The larger the value, the better the overall performance in terms of motion and energy saving. It is used for GAE advantage estimation and network loss calculation.
[0135] In this embodiment, the weight coefficients α, β, γ, and δ correspond to the adjustment weights of each sub-item, used to control the degree of influence of each item on the total reward: α is the weight of the linear velocity + angular velocity tracking item, α is greater than 0, positively amplifying the speed tracking benefit; β is the weight of the resonance recognition reward, β is greater than 0, prioritizing the excitation of the step frequency to match the spring resonance frequency, which is the core optimization target; γ is the weight of the joint torque penalty, γ is less than 0, the greater the torque, the greater the decrease in the total reward, constraining the motor output; δ is the weight of the motion stability penalty, δ is less than 0, the greater the body tilt, the greater the decrease in the total reward, preventing instability.
[0136] Specifically, the method for determining the weight coefficients of α, β, γ, and δ is as follows: First, multiple sets of different weight combinations are built to conduct simulation comparison experiments. With the terrain, gait, and network structure fixed, the values of α, β, γ, and δ are adjusted in sequence, and the four core indicators of robot speed tracking error, step frequency and instantaneous natural frequency deviation of spring, average output torque of motor, and body pitch and roll offset are statistically analyzed. With resonance matching accuracy and motor power consumption reduction as the core optimization objectives, while taking into account the accuracy of travel and steering and the stability of body movement, the weight combination that can simultaneously achieve accurate speed tracking, full utilization of resonance, effective suppression of motor torque, and controllable body tilt is selected. Finally, it is determined that α and β are positive to enhance the positive excitation of speed and resonance, and γ and δ are negative to impose penalty constraints on torque and body tilt.
[0137] In this embodiment, r_lv is the linear velocity tracking reward, which is a positive component. It is calculated based on the deviation between the actual x and y travel speeds of the torso and the desired speed. The smaller the deviation, the higher the value, ensuring that the robot accurately follows forward and lateral movement commands.
[0138] Specifically, the linear velocity reward r_lv incentivizes the policy to improve its ability to track the desired linear velocity on the horizontal plane (in the x and y directions), and the formula is as follows:
[0139] ,in, It is the expected linear velocity. It is the current linear velocity, which is based on the body's center of mass.
[0140] In this embodiment, r_av is the yaw rate tracking reward, which is a positive component. It is calculated based on the difference between the actual turning angular velocity of the fuselage and the expected yaw angular velocity. The smaller the difference, the higher the reward, thus achieving precise steering control.
[0141] Specifically, the angular velocity reward r_av improves the strategy's ability to track the desired yaw angular velocity, and the formula is:
[0142] ,in, It is the expected angular velocity. It is the current angular velocity, which is based on the body's center of mass.
[0143] In this embodiment, r_τ is a joint torque penalty term, which is a penalty component. Its value increases as the output torque of each joint motor increases. After being multiplied by the negative weight γ, it deducts from the total reward, thereby suppressing the peak torque output of the motor.
[0144] Specifically, the joint torque penalty r_τ encourages the strategy to output the lowest possible joint torque, thereby improving energy efficiency. The formula is as follows:
[0145] , where τ i It is the torque of the i-th joint calculated from the current.
[0146] In this embodiment, r_r is the resonance recognition reward, which is the core positive component. It compares the real-time motion step frequency with the instantaneous natural frequency of the spring. The closer the two are, the higher the value, which specifically guides the controller to use spring resonance to reduce motor power consumption. This represents a constant between 0 and 1. During training, it gradually changes from 0 to 1, indicating that these penalty terms are gradually introduced into the training process.
[0147] Specifically, the resonance recognition reward r_r motivates the strategy to improve its tracking ability of the joint's equivalent natural frequency, as shown in the formula:
[0148] ,in, It is the equivalent natural frequency of the i-th joint calculated from simulation environment variables. The frequency estimate output by the policy network for the i-th joint. This represents a constant between 0 and 1. During training, it gradually changes from 0 to 1, indicating that these penalty terms are gradually introduced into the training process.
[0149] In this embodiment, r_stab is a motion stability penalty term, which is a penalty sub-term calculated based on the offset of the fuselage pitch angle and roll angle relative to the horizontal reference. The more severe the tilt, the larger the value of this term. After negative weighting δ to reduce the total reward, the fuselage tilting and gait imbalance are avoided.
[0150] In this embodiment, the fuselage tilt penalty term, r_stab, is a constraint term that takes the fuselage pitch and roll offset as input and is a key penalty module to ensure motion stability.
[0151] As can be seen from the above, this embodiment achieves multi-objective quantitative control by designing a linear composite reward function with weighted coefficients and distinguishing between incentive and penalty terms through positive and negative weights. The resonance matching reward weight is a positive high weight, which prioritizes driving the algorithm to converge to the resonance interval; the torque and stability terms are set with negative weights to continuously suppress large torque and body tilt, preventing conflicts between energy saving and motion stability. Each weight can be flexibly adjusted according to the robot model, spring stiffness, and motion scenario to adapt to different working conditions such as high-speed walking and low-speed loading; the formulaic quantitative reward facilitates batch parameter tuning in simulation, accelerates network convergence, stably guides the controller to output the optimal step frequency that matches the spring resonance frequency, and continuously reduces the average output torque of the motor. While ensuring the stability of the robot body and accurate following of speed commands, it maximizes the reduction of the overall machine's operating power consumption.
[0152] In one embodiment of this application, foot force sensing data, proprioceptive sensing data, and environmental implicit representation vectors are input into a pre-trained spring natural frequency evaluation network to estimate and output the current instantaneous natural frequency of the passive elastic element at the joint online, and the network further includes:
[0153] The baseline value for the current instantaneous natural frequency of the online spring natural frequency estimation network includes:
[0154] Obtain the mass parameters, center of mass position parameters, and inertia tensor parameters of all links driven by the joint;
[0155] Obtain the inertia parameters of the motor rotor and the gear reduction ratio parameters;
[0156] By using the parallel axis theorem to convert all moments of inertia onto the joint rotation axis and summing them, the equivalent moment of inertia can be obtained.
[0157] The natural frequency reference value is calculated based on the equivalent moment of inertia.
[0158] The natural frequency reference value is used to supervise or constrain the output of the spring natural frequency evaluation network.
[0159] In this embodiment, the passive elastic element at the joint is a torsion spring installed in parallel at each knee joint of the leg, which relies on deformation to realize the mutual conversion of kinetic energy and elastic potential energy.
[0160] In this embodiment, the current instantaneous natural frequency is the real-time equivalent oscillation frequency of the spring-link system under the current terrain, leg posture, and load, output by network inference. The natural frequency reference value is the theoretical standard resonant frequency calculated using the robot's mechanical structure parameters and rigid body dynamics formulas, serving as a physical truth reference.
[0161] In this embodiment, the mass parameter, center of mass position parameter, and inertia tensor parameter of the link are inherent mechanical parameters of the robot's thigh, calf, and other leg components: the mass is the weight of the link; the center of mass position is the link's center of gravity relative to the joint coordinates; and the inertia tensor characterizes the rotational inertia of the link about its own central axis.
[0162] In this embodiment, the inertia parameter of the motor rotor is the rotational inertia of the motor output shaft itself, which will be included in the calculation of the total rotational inertia of the joint. The gear reduction ratio parameter is the transmission ratio of the joint reducer, and the inertia on the motor side needs to be converted to the joint rotation shaft according to the reduction ratio.
[0163] In this embodiment, the parallel axis theorem is a fundamental theorem of rigid body mechanics, used to uniformly convert the moment of inertia of the link center that deviates from the joint rotation axis to the knee joint rotation center.
[0164] In this embodiment, the equivalent moment of inertia is the sum of the converted moments of inertia of all connecting rods and the motor rotor inertia after reduction ratio conversion, which are then uniformly equivalent to the total moment of inertia of the joint shaft.
[0165] In this embodiment, supervising or constraining the network output includes two application methods: during the training phase, the benchmark value is used as a label to supervise the network learning; during the deployment and inference phase, the benchmark value is used as an upper and lower limit constraint, and if the network prediction frequency deviates too much from the benchmark, it is truncated and corrected to prevent prediction anomalies.
[0166] As can be seen from the above, this embodiment calculates the equivalent moment of inertia online and derives the natural frequency reference value. By integrating all inertia parameters of the connecting rod, motor rotor, and reduction mechanism through the parallel axis theorem, a frequency supervision reference with complete physical basis is constructed, providing real-time constraint correction for the spring frequency evaluation network output. This dynamically compensates for equivalent inertia fluctuations caused by load changes and leg posture alterations, solving the problem of pure data-driven network prediction drift. The physical reference participates in network inference supervision in real time, reducing frequency estimation errors and preventing a surge in power consumption caused by step frequency deviating from the resonance range. No additional calibration equipment is required; the correction reference can be generated in real time using only the robot's inherent structural parameters, balancing physical reliability and real-time computational efficiency, improving resonance matching accuracy, and continuously optimizing motor energy consumption.
[0167] In one embodiment of this application, a control method for optimizing the power consumption of a quadruped robot motor using spring resonance characteristics further includes strategy deployment:
[0168] Export the trained external perception encoder, state evaluation network, spring natural frequency evaluation network, and main network as policy files;
[0169] The strategy file is loaded into the onboard computing unit of the quadruped robot and runs as a real-time controller;
[0170] The real-time controller executes steps S101 to S106 cyclically at a preset control frequency.
[0171] In this embodiment, strategy deployment is the process of porting and loading the network model to the real quadruped robot hardware after the reinforcement learning model has converged during offline training, and combining it with the local real-time control program.
[0172] In this embodiment, the external perception encoder, state evaluation network, spring natural frequency evaluation network, and main network after training are four types of sub-networks that are integrated to form the Actor policy network. After multi-terrain simulation iterative convergence, the weight parameters have been optimized to adapt to the resonant energy-saving gait; the Critic network used only for training is not included.
[0173] In this embodiment, exporting as a strategy file involves uniformly encapsulating and lightweighting the weights and network structures of the four sub-networks to generate a model file that can be read by embedded devices (such as ONNX, pt, and TensorRT models), removing training-related gradients and optimizer parameters, and reducing storage usage.
[0174] In this embodiment, the onboard computing unit is an embedded AI processor, such as Jetson Nano or Jetson Xavier, mounted inside the quadruped robot's body. It has real-time neural network inference capabilities, does not rely on the cloud, and can complete all algorithm calculations locally.
[0175] In this embodiment, the real-time controller is an embedded control program that runs after loading the strategy file. It independently completes the full control logic of S101-S106 and periodically outputs motor commands to drive leg movement.
[0176] In this embodiment, the preset control frequency is the refresh frequency of the controller's closed-loop cyclic execution algorithm, in Hz. The higher the frequency, the smoother the gait control. Setting it to 100Hz-200Hz meets the robot's dynamic movement requirements.
[0177] The preset control frequency is determined by considering the neural network inference time of the onboard computing unit, the response bandwidth of the robot's servo motor, and the dynamic motion response requirements of the legs. First, the total time for the inference, data transmission, and command issuance of the entire set of algorithms S101-S106 for a single wheel is measured. Based on this, a 30% redundancy cycle is reserved to avoid control cycle timeout. At the same time, the standard control bandwidth of the servo drive is matched to take into account the dynamic working conditions such as the rapid leg swing and landing impact of the quadruped robot. Finally, a stable preset control frequency without frame loss is determined.
[0178] In this embodiment, the cyclic execution of steps S101 to S106 means that each control cycle sequentially completes the sensing acquisition, environmental encoding, state fusion, instantaneous natural frequency online estimation, motor command generation, and motor drive execution, continuously adjusting the step frequency to match the spring resonance frequency.
[0179] As can be seen from the above, this embodiment exports the entire Actor network as an independent policy file after training and directly loads it into the onboard computing unit to form a dedicated real-time controller. This allows it to run independently of the simulation environment, making the deployment process simple and efficient. The controller cyclically executes the entire process of perception, encoding, frequency estimation, and instruction generation according to a fixed high-frequency cycle, ensuring stable and controllable control timing and guaranteeing that the step frequency follows the instantaneous natural frequency of the spring in real time. The inference link has no offline computation or cloud interaction; all computations are performed locally onboard, making it suitable for outdoor scenarios without communication. The lightweight policy file consumes little storage resources, adapting to embedded low-computing-power hardware. Resonance energy-saving control is achieved without upgrading the onboard hardware, expanding the algorithm's applicable scenarios.
[0180] Corresponding to the control method for optimizing the power consumption of quadruped robot motors using the spring resonance characteristics described in the above embodiment, Figure 4 This is a structural block diagram of a control system for optimizing the power consumption of a quadruped robot motor using the resonant characteristics of a spring, provided as an embodiment of this application. For ease of explanation, only the parts relevant to the embodiment of this application are shown. References Figure 4 The control system 20, which utilizes the resonant characteristics of springs to optimize the power consumption of the quadruped robot motor, includes: a perception data module 21, an environmental terrain module 22, a state evaluation module 23, a natural frequency module 24, a motor command module 25, and a tracking control module 26.
[0181] Among them, the perception data module 21 is used to acquire the perception data of the quadruped robot, including body perception data, external environment perception data and foot force perception data.
[0182] The environmental terrain module 22 is used to input external environmental perception data into the external perception encoder to obtain the implicit environmental representation vector;
[0183] The state evaluation module 23 is used to input the ontology perception data and the implicit representation vector of the environment into the pre-trained state evaluation network to obtain the implicit representation vector of the overall characteristics of the robot.
[0184] The natural frequency module 24 is used to input foot force sensing data, body sensing data and environmental implicit expression vector into a pre-trained spring natural frequency evaluation network to estimate and output the current instantaneous natural frequency of the passive elastic element at the joint online.
[0185] The motor instruction module 25 is used to input the current instantaneous natural frequency, the implicit state representation vector, and the ontology perception data into the pre-trained main network to generate motor action instructions;
[0186] The tracking control module 26 is used to control the operation of the quadruped robot's motor according to the motor action command, so that the quadruped robot's step frequency tracks the current instantaneous natural frequency and reduces motor power consumption.
[0187] Each leg of the quadruped robot has a spring connected in parallel at its knee joint.
[0188] See Figure 5 This is a hardware architecture diagram of a quadruped robot, including: the quadruped robot body with elastic springs installed in parallel at its knee joints; an AI computing processor located in the robot's torso to run a hierarchical policy network and generate motor motion commands; an IMU located in the robot's torso and joint encoders located in each joint to collect body posture, velocity, joint angles, and angular velocities; current sensors located in each motor drive circuit to collect motor current; six-dimensional force sensors located at the foot tips to collect foot contact forces; a depth camera and LiDAR located in the robot's torso to collect external environment perception data; and motor controllers and drive motors located in each joint to receive motion commands from the AI computing processor and drive the robot's movement. The AI computing processor communicates with the IMU, joint encoders, current sensors, foot force sensors, depth camera, LiDAR, and motor controllers to form a complete closed loop of perception-decision-control. In specific implementation, the AI computing processor has communication connections with the IMU, joint encoder, current sensor, foot six-dimensional force sensor, depth camera and LiDAR. The AI computing processor is deployed with a pre-trained hierarchical policy network, which includes an external perception encoder, a state evaluation network, a spring natural frequency evaluation network and a main network. It can execute the implementation described in any embodiment of the control method for optimizing the power consumption of quadruped robot motors using spring resonance characteristics provided in the embodiments of this application, and can also execute the implementation of the electronic device described in the embodiments of this application.
[0189] See Figure 6 , Figure 6 This is a schematic block diagram of an electronic device provided according to an embodiment of this application. Figure 6The electronic device 300 in this embodiment may include one or more processors 301, one or more input devices 302, one or more output devices 303, and one or more memories 304. The processors 301, input devices 302, output devices 303, and memories 304 communicate with each other via a communication bus 305. The memories 304 store computer programs, including program instructions. The processors 301 execute the program instructions stored in the memories 304. Specifically, the processors 301 are configured to invoke the program instructions to perform the functions of the modules in the aforementioned device embodiments, for example... Figure 4 The functions of the sensing data module 21, environmental terrain module 22, state assessment module 23, natural frequency module 24, motor command module 25, and tracking control module 26 are shown.
[0190] It should be understood that, in the embodiments of this application, the processor 301 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.
[0191] Input device 302 may include a touchpad, a fingerprint sensor (for collecting the user's fingerprint information and fingerprint orientation information), a microphone, etc., and output device 303 may include a display (LCD, etc.), a speaker, etc.
[0192] The memory 304 may include read-only memory and random access memory, and provides instructions and data to the processor 301. A portion of the memory 304 may also include non-volatile random access memory. For example, the memory 304 may also store device type information.
[0193] In specific implementations, the processor 301, input device 302, and output device 303 described in the embodiments of this application can execute the implementation methods described in any embodiment of the control method for optimizing the power consumption of a quadruped robot motor using the spring resonance characteristics provided in the embodiments of this application, or they can execute the implementation methods of the electronic devices described in the embodiments of this application, which will not be repeated here.
[0194] In another embodiment of this application, a computer-readable storage medium is provided. This computer-readable storage medium stores a computer program, which includes program instructions. When executed by a processor, the program instructions implement all or part of the processes in the methods described above. Alternatively, the computer program can instruct related hardware to complete the process. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include any entity or device capable of carrying computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.
[0195] The computer-readable storage medium can be an internal storage unit of the electronic device in any of the foregoing embodiments, such as a hard disk or memory of the electronic device. The computer-readable storage medium can also be an external storage device of the electronic device, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the electronic device. Furthermore, the computer-readable storage medium can include both internal and external storage units of the electronic device. The computer-readable storage medium is used to store computer programs and other programs and data required by the electronic device. The computer-readable storage medium can also be used to temporarily store data that has been output or will be output.
[0196] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this application.
[0197] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the electronic devices and units described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.
[0198] In the several embodiments provided in this application, it should be understood that the disclosed electronic devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces or units, or it may be an electrical, mechanical, or other form of connection.
[0199] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of the embodiments of this application, depending on actual needs.
[0200] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.
[0201] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A control method for optimizing the power consumption of a quadruped robot motor using the resonance characteristics of a spring, characterized in that, include: S101: Acquire the perception data of the quadruped robot, including body perception data, external environment perception data, and foot force perception data; S102: Input the external environment perception data into the external perception encoder to obtain the implicit environmental representation vector; S103: Input the ontology perception data and the environment implicit representation vector into the pre-trained state evaluation network to obtain the state implicit representation vector of the robot's overall characteristics; S104: Input the foot force sensing data, the body sensing data and the environment implicit expression vector into the pre-trained spring natural frequency evaluation network, estimate and output the current instantaneous natural frequency of the passive elastic element at the joint online; S105: Input the current instantaneous inherent frequency, the state implicit expression vector, and the ontology perception data into the pre-trained main network to generate motor action commands; S106: Control the operation of the quadruped robot's motor according to the motor action command, so that the quadruped robot's step frequency tracks the current instantaneous natural frequency, thereby reducing motor power consumption; The quadruped robot has elastic springs connected in parallel at the knee joint of each leg.
2. The control method for optimizing the power consumption of a quadruped robot motor using the resonant characteristics of a spring, as described in claim 1, is characterized in that... The step of inputting the external environment perception data into the external perception encoder to obtain the implicit environmental representation vector includes: The external perception encoder uses a convolutional neural network to extract features and compress dimensions from the external environment perception data. The environmental implicit representation vector is used to characterize one or more of the following: terrain type, terrain roughness, ground stiffness characteristics, and ground damping characteristics.
3. The control method for optimizing the power consumption of a quadruped robot motor using the resonant characteristics of a spring, as described in claim 1, is characterized in that... The pre-trained spring natural frequency evaluation network is a recurrent neural network, including: The recurrent neural network includes LSTM units or GRU units, which model the temporal dynamic characteristics of the input data to track and output the current instantaneous natural frequency online. During the training phase, the equivalent rotational inertia of the joints is calculated in advance based on the mass, center of mass, inertia, motor rotor inertia, and reduction ratio of the robot links using the parallel axis theorem. The theoretical natural frequency is obtained through the spring stiffness coefficient and used as the training benchmark label for the spring natural frequency evaluation network.
4. The control method for optimizing the power consumption of a quadruped robot motor using the resonant characteristics of a spring, as described in claim 1, is characterized in that... Also includes: The external perception encoder, the state evaluation network, the spring natural frequency evaluation network, and the main network constitute the Actor policy network, which is trained using the reinforcement learning Actor-Critic framework. During training, the input to the Critic network includes the perception data and privileged information, which includes one or more of the following: ground friction coefficient, ground stiffness parameter, leg collision status, and flight duration. After training, only the Actor policy network is exported for deployment, and the Critic network does not participate in the inference computation during the deployment phase.
5. The control method for optimizing the power consumption of a quadruped robot motor using the resonant characteristics of a spring, as described in claim 4, is characterized in that... The training process of the reinforcement learning Actor-Critic framework includes: S201: Construct a dynamic model of the quadruped robot in a simulation environment and set up various training terrains; S202: The perception data is input into the Actor policy network, and the Actor policy network outputs action commands and applies them to the dynamic model of the quadruped robot in the simulation environment, driving the dynamic model of the quadruped robot to perform actions; S203: Input the perceived data and the privileged information into the Critic network, and output the value estimate of the current state; S204: Calculate the reward value based on the state change after the quadruped robot performs the action according to the dynamic model of the quadruped robot, and use the GAE algorithm to calculate the advantage estimate; S205: Calculate the policy loss and value function loss based on the advantage estimation, and update the network parameters of the Actor policy network and the Critic network through the optimizer in a backpropagation manner; S206: Repeat steps S202 to S205 until the network converges.
6. The control method for optimizing the power consumption of a quadruped robot motor using the resonant characteristics of a spring, as described in claim 5, is characterized in that... The calculation of the reward value based on the state change after the quadruped robot performs an action according to its dynamic model includes: The reward value is calculated by a reward function, which includes multiple of the following reward items: linear velocity tracking reward item, angular velocity tracking reward item, resonance recognition reward item, joint torque penalty item, and motion stability penalty item; The linear velocity tracking reward is used to incentivize tracking of the desired linear velocity in the horizontal direction based on the deviation between the robot's actual linear velocity and the desired linear velocity. The angular velocity tracking reward is used to incentivize tracking of the desired yaw angular velocity based on the deviation between the robot's actual angular velocity and the desired angular velocity. The resonance recognition reward is used to incentivize the movement cadence to track the current instantaneous natural frequency based on the deviation between the current instantaneous natural frequency and the current movement cadence. The joint torque penalty term is used to constrain the motor output torque based on the amplitude of the output torque of each joint motor. The motion stability penalty is used to penalize the fuselage tilt angle when it exceeds the allowable range, based on the deviation of the fuselage roll angle and pitch angle.
7. The control method for optimizing the power consumption of a quadruped robot motor using the resonant characteristics of a spring, as described in claim 6, is characterized in that... The reward function includes: The reward function R = α·(r_lv+r_av)+β·r_r+γ·r_τ+δ·r_stab; Wherein, α, β, γ, and δ are the weight coefficients corresponding to each reward item, α > 0, β > 0, γ < 0, and δ < 0; r_lv is the linear velocity tracking reward, used to constrain the robot's torso's x and y direction travel speed to match the desired command; r_av is the yaw rate tracking reward, used to constrain the body's turning rate to match the desired command; r_r is the resonance recognition reward, used to guide the frequency estimate output by the network to match the real-time equivalent natural frequency of the joint; r_τ is the joint torque penalty term, used to suppress the motor's output of large torque to reduce power consumption; r_stab is the motion stability penalty term; the reward function is supplemented with a body tilt penalty term to ensure the robot's motion stability.
8. The control method for optimizing the power consumption of a quadruped robot motor using the resonant characteristics of a spring, as described in claim 1, is characterized in that... The step of inputting the foot force sensing data, the body sensing data, and the environmental implicit representation vector into a pre-trained spring natural frequency evaluation network to estimate and output the current instantaneous natural frequency of the passive elastic element at the joint online further includes: The online estimation of the reference value of the current instantaneous natural frequency of the spring natural frequency evaluation network includes: Obtain the mass parameters, center of mass position parameters, and inertia tensor parameters of all links driven by the joint; Obtain the inertia parameters of the motor rotor and the gear reduction ratio parameters; By using the parallel axis theorem to convert all moments of inertia onto the joint rotation axis and summing them, the equivalent moment of inertia can be obtained. The natural frequency reference value is calculated based on the equivalent moment of inertia. The natural frequency reference value is used to supervise or constrain the output of the spring natural frequency evaluation network.
9. The control method for optimizing the power consumption of a quadruped robot motor using the resonant characteristics of a spring, as described in claim 1, is characterized in that... This also includes strategy deployment: The trained external perception encoder, the state evaluation network, the spring natural frequency evaluation network, and the main network are exported as a policy file. The strategy file is loaded into the onboard computing unit of the quadruped robot and runs as a real-time controller; The real-time controller therein executes steps S101 to S106 cyclically at a preset control frequency.
10. A control system for optimizing the power consumption of a quadruped robot motor by utilizing the resonance characteristics of a spring, characterized in that, include: The perception data module is used to acquire the perception data of the quadruped robot, including body perception data, external environment perception data, and foot force perception data. The environment and terrain module is used to input the external environment perception data into the external perception encoder to obtain the implicit environmental representation vector; The state evaluation module is used to input the ontology perception data and the environment implicit expression vector into a pre-trained state evaluation network to obtain the state implicit expression vector of the robot's overall characteristics. The natural frequency module is used to input the foot force sensing data, the body sensing data and the environmental implicit expression vector into a pre-trained spring natural frequency evaluation network, and to estimate and output the current instantaneous natural frequency of the passive elastic element at the joint online. The motor instruction module is used to input the current instantaneous natural frequency, the implicit state representation vector, and the ontology perception data into the pre-trained main network to generate motor action instructions; The tracking control module is used to control the operation of the motor of the quadruped robot according to the motor action command, so that the step frequency of the quadruped robot tracks the current instantaneous natural frequency and reduces the power consumption of the motor. The quadruped robot has elastic springs connected in parallel at the knee joint of each leg.
