Amphibious robot energy efficiency optimization control method for amphibious cross-domain operation

Abstract

The invention discloses an amphibious robot energy efficiency optimization control method for amphibious cross-domain operation, and belongs to the technical field of robot control, and the method comprises the steps: obtaining multi-modal sensor data of an amphibious robot, carrying out the feature extraction of the multi-modal sensor data, and generating a coupling feature vector; and the stability constraint trained neural network generates a dynamic value function, predicts the time-varying constraint in a future time domain through a system evolution model based on the dynamic value function and the coupling feature vector, and generates a correction value to adjust the structure parameter of the dynamic value function. Performing real-time nonlinear optimization on the corrected dynamic value function, and solving an optimal control instruction; collecting actual response data of the system, and updating a model used for feature extraction and prediction on line by using the actual response data of the system; on the premise that the system stability is guaranteed, the environmental adaptability and the overall energy efficiency of the robot in the water and land cross-domain process are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of robot control technology, and in particular to an energy efficiency optimization control method for amphibious robots operating across land and water. Background Technology

[0002] Amphibious robots are a type of specialized robot capable of performing tasks in various environments, including land and water. Their control system is the core of their autonomous, efficient, and stable operation. The main task of the control system is to generate appropriate control commands to drive actuators based on task requirements and environmental information, thereby completing tasks such as path tracking, attitude maintenance, and target acquisition. These control systems typically fall under the category of programmable control or adaptive control systems, and their performance directly determines the application range and operational efficiency of the amphibious robot.

[0003] In existing technologies, control methods for amphibious robots typically employ traditional control architectures, such as linear quadratic regulators based on accurate models, and some advanced model predictive control methods. These methods mostly rely on a pre-defined cost function or performance metric that remains constant throughout the task. This cost function is usually a weighted sum of multiple objectives, such as tracking error and energy consumption. The controller's task is to find the control input that minimizes this cost function at each time step.

[0004] However, the aforementioned existing technical solutions have significant technical shortcomings when dealing with task scenarios involving drastic and nonlinear changes in physical characteristics, such as those spanning land and water. Their control objective, i.e., the cost function, remains fixed, making them unable to adapt to dynamic environmental changes. Weight parameters optimized for terrestrial environments may lead to low energy efficiency or sluggish control response in underwater environments, and vice versa. Furthermore, these methods are typically optimized based on the current state, lacking the ability to predict changes in key future constraints. For example, they cannot foresee the loss of GPS signal due to impending water entry, thus failing to adjust control strategies in advance to address this change. This often results in a sharp performance drop or even control instability during environmental transitions. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides an energy efficiency optimization control method for amphibious robots operating across land and water. This method employs a closed-loop collaborative control approach that involves real-time construction of a dynamic value function, forward-looking correction of that function, and online optimization of the function. This significantly improves the robot's environmental adaptability and overall energy efficiency during cross-land and water operations while ensuring system stability.

[0006] The above objectives can be achieved through the following approach:

[0007] An energy efficiency optimization control method for amphibious robots operating across land and water, the method comprising:

[0008] Acquire multimodal sensor data from the amphibious robot, extract features from the multimodal sensor data, and generate coupled feature vectors;

[0009] Based on coupled feature vectors, a neural network trained with stability constraints generates a dynamic value function, whose structural parameters automatically adjust as the environment changes.

[0010] Based on the dynamic value function and the coupled feature vector, the time-varying constraints in the future time domain are predicted by the system evolution model, and a correction amount is generated to adjust the structural parameters of the dynamic value function to obtain the corrected dynamic value function.

[0011] Real-time nonlinear optimization is performed on the corrected dynamic value function to solve for the optimal control command;

[0012] The optimal control command is executed, and the actual system response data is collected. The model used for feature extraction and prediction is then updated online using the actual system response data.

[0013] Optionally, the step of extracting features from multimodal sensor data and generating a coupled feature vector includes: separating internal state data and external environment data from the multimodal sensor data; performing coupled analysis on the internal state data and the external environment data to extract deep features reflecting the overall operating quality of the system; and generating the coupled feature vector based on the deep features.

[0014] Optionally, the step of generating the dynamic value function includes: inputting the coupled feature vector into the neural network to obtain the structural parameters of the dynamic value function; and constructing the dynamic value function based on the structural parameters, wherein the dynamic value function includes penalty weights on the higher-order derivatives of the control quantity.

[0015] Furthermore, when constructing the dynamic value function, Lyapunov stability constraints are embedded to ensure that the dynamic value function can be guided to a stable controller.

[0016] Optionally, the step of obtaining the corrected dynamic value function includes: using the dynamic value function and the coupled feature vector as initial conditions, inputting them into the system evolution model to obtain the time-varying constraints in the future time domain; based on the time-varying constraints in the future time domain, deriving the correction amount used to adjust the structural parameters; and applying the correction amount to the structural parameters of the dynamic value function to obtain the corrected dynamic value function.

[0017] Optionally, the step of solving for the optimal control command includes: using the corrected dynamic value function as the objective function to construct a nonlinear model predictive control optimization problem; performing sensitivity analysis based on the optimal solution at the previous time step to generate the initial iteration point of the current optimization problem; and iteratively solving from the initial iteration point to obtain the optimal control command.

[0018] Optionally, the step of updating the model used for feature extraction and prediction online includes: comparing the actual response data of the system with the predicted response when the optimal control command is issued to generate a prediction error; and using the prediction error to fine-tune online the feature extraction model used to generate the coupled feature vector and the system evolution model used to predict the time-varying constraints.

[0019] Optionally, the deep features include an estimated medium viscosity coefficient, actuator load rate variance, and sensor information entropy; the estimated medium viscosity coefficient is used to characterize the resistance characteristics of the current environment, the actuator load rate variance is used to quantify the degree of fluctuation in actuator load, and the sensor information entropy is used to assess the uncertainty of sensor data.

[0020] Optionally, the penalty weight for the higher-order derivative of the control quantity is automatically increased or decreased according to the chaotic or stable environment represented by the coupled feature vector; the penalty weight is increased in a chaotic environment to suppress jitter, and the penalty weight is decreased in a stable environment to focus on energy consumption optimization.

[0021] Optionally, the time-varying constraints in the future time domain include actuator saturation boundary, sensor availability, and energy recovery potential; the actuator saturation boundary is used to warn of the upper limit of control capability, the sensor availability is used to warn of the loss of sensing information, and the energy recovery potential is used to identify the external energy that can be used in the future.

[0022] Optionally, the step of constructing a nonlinear model predictive control optimization problem includes dynamically constructing a set of constraints for the real-time optimization problem based on the structural parameters of the coupled feature vector and the corrected dynamic value function; when the control smoothness penalty weight in the structural parameters is greater than a first threshold, tightening the actuator peak torque constraint from the rated torque value to a lower robust torque value; and when the estimated value of the medium viscosity coefficient in the coupled feature vector indicates a medium switching, adding an idling constraint to limit the thruster idling speed to the set of constraints.

[0023] Compared with the prior art, the present invention has the following advantages:

[0024] This invention achieves a synergistic effect by constructing a dynamic value function, performing forward-looking correction, and combining it with real-time optimization. This results in significant benefits, improving the robot's environmental adaptability and operational stability. By sensing the environment and the robot's state in real time and dynamically constructing the value function, the core objective of the control system is no longer rigid. It can intelligently adapt to the drastic changes in the physical characteristics of water and land media, automatically prioritizing stability in turbulent or transitional regions. This suppresses control jitter and instability risks, ensuring smooth and safe cross-domain operations.

[0025] This invention introduces a prediction and correction mechanism for future time-varying constraints, making control decisions more forward-looking. The system can adjust the current control objectives in advance before entering dangerous or adverse operating conditions, preparing for the upcoming challenges and avoiding the creation of hidden dangers for the future system state by pursuing the optimal state at the current moment. This significantly improves the global energy efficiency and success rate of the entire task profile and realizes the upgrade of the control strategy from local optimum to global optimum.

[0026] This invention ensures the engineering feasibility and continuous evolution capability of advanced theories through tightly coupled closed-loop learning and real-time optimization. The highly efficient optimizer, specifically designed for dynamic objective functions, ensures real-time output of complex decisions, transforming perception and planning into precise physical actions. Meanwhile, the feedback of execution results continuously optimizes the internal model, enabling the system to learn online and improve itself. This enhances the robustness to model uncertainties and unknown environmental disturbances, allowing the robot's overall performance to continuously improve with the accumulation of operational experience. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0029] Figure 2 This is a schematic diagram of the dynamic value function and correction mechanism of the present invention;

[0030] Figure 3 This is a schematic diagram of the nonlinear model predictive control of the present invention;

[0031] Figure 4 This is a schematic diagram comparing the effects of the conventional method and the method of the present invention in an embodiment of the present invention. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] Reference Figure 1 As shown, one embodiment of the present invention proposes an energy efficiency optimization control method for amphibious robots operating across land and water, the method comprising:

[0034] Acquire multimodal sensor data from the amphibious robot, extract features from the multimodal sensor data, and generate coupled feature vectors;

[0035] Based on coupled feature vectors, a neural network trained with stability constraints generates a dynamic value function, whose structural parameters automatically adjust as the environment changes.

[0036] Based on the dynamic value function and coupled eigenvectors, the system evolution model predicts the time-varying constraints in the future time domain and generates correction values ​​to adjust the structural parameters of the dynamic value function, thus obtaining the corrected dynamic value function.

[0037] Real-time nonlinear optimization is performed on the corrected dynamic value function to solve for the optimal control command;

[0038] Execute optimal control commands and collect actual system response data, then use the actual system response data to update the model used for feature extraction and prediction online.

[0039] Specifically, through five interdependent and sequentially executed steps, the amphibious robot achieves high adaptability and forward-looking optimization in complex cross-domain environments. By deeply fusing the robot's internal and external multimodal perception data, a coupling feature vector is extracted to refine the current state of the robot-environment coupled system. Using a pre-trained neural network, this feature vector is mapped in real time to a dynamically adjusted value function, which formally defines the current optimal control objective. Next, the method introduces a prediction mechanism to predict potential future changes in system constraints based on the current state and control tendency, and uses this to correct the current value function, injecting it with foresight. An efficient nonlinear optimization algorithm is used to solve this time-varying value function with look-ahead correction in real time, generating optimal physical control commands. The actual execution effect of the control commands is used as feedback to continuously update and optimize the internal feature extraction and prediction models online, forming a complete, self-improving intelligent closed loop of perception and optimization feedback.

[0040] By constructing a dynamic value function in real time, the robot can intelligently adjust its control priority according to environmental changes. In complex environments, it prioritizes stability, while in stable environments, it prioritizes energy efficiency, thus overcoming the limitation of traditional control methods that struggle to balance stability and energy efficiency. The forward-looking correction mechanism enables the robot to anticipate and mitigate future risks, avoiding control failures or energy waste caused by short-sighted decisions, achieving a leap from local optima to global optima. Efficient online optimization and closed-loop learning capabilities ensure the real-time nature and accuracy of control decisions, enabling the robot to continuously learn and adapt in unknown and changing environments. Ultimately, this method results in amphibious robots exhibiting higher environmental adaptability, better energy utilization efficiency, stronger robustness, and smoother task transition capabilities.

[0041] Optionally, the steps of extracting features from multimodal sensor data and generating coupled feature vectors include: separating internal state data and external environment data from multimodal sensor data; performing coupled analysis on the internal state data and external environment data to extract deep features reflecting the overall operating quality of the system; and generating coupled feature vectors based on the deep features.

[0042] Specifically, the system acquires real-time operational data of the amphibious robot through onboard multimodal sensors. This data is divided into internal state data and external environment data. Internal state data includes information such as actuator load rate, battery voltage, angular velocity, and acceleration from motor current sensors, battery management system, and inertial measurement unit, reflecting the robot's own operating conditions. External environment data includes information such as depth, current velocity, and location from water depth sensors, flow meters, or GPS, describing the robot's operating environment.

[0043] To extract deep features reflecting the overall operational quality of the system, the system performs coupled analysis of acquired internal state data and external environmental data. For example, by combining the real-time torque output of the actuators (internal state data) with the robot's motion speed measured by the inertial measurement unit (IMU) (internal state data), the system dynamically estimates the viscosity coefficient of the medium in the current working environment based on a preset fluid dynamics model. Furthermore, the system collects the actuator load rate sequence within a sliding time window and calculates its variance to obtain the actuator load rate variance. This feature is used to quantify the stability of the control load. Simultaneously, to assess the uncertainty of perceived information and the complexity of the environment, the system discretizes the output signals of key sensors, such as the IMU, within a time window and calculates their information entropy as the sensor information entropy. The specific calculation is as follows:

[0044] ,

[0045] in, Represents the sensor signal sequence. It is a specific value or range of the signal after it has been discretized. It is the probability of the value or interval appearing within the time window, obtained by statistically analyzing its frequency.

[0046] The estimated viscosity coefficient of the medium, the variance of the actuator load rate, and the information entropy of the sensors, obtained through the above coupling analysis, are combined into a numerical, low-dimensional coupling feature vector between the environment and the robot. This vector has a compact structure and high information density, which can comprehensively reflect the dynamic characteristics of the robot's interaction with the environment at the current moment. It also serves as the sole input for the subsequent generation of the dynamic value function, ensuring that the basis for subsequent decisions is the system state that has been deeply refined and characterized.

[0047] By performing coupled analysis and deep feature extraction on the raw multimodal sensor data, the transformation from massive heterogeneous raw perception information to low-dimensional information condensation of system state representation was realized. The internal working conditions of the robot and the characteristics of the external environment were organically integrated into a unified feature vector, avoiding the subsequent processing modules from directly facing the complexity and high dimensionality of the raw data. This provides a solid and reliable data foundation for the entire control system to achieve accurate perception and adaptive decision-making in complex dynamic environments, thereby significantly improving the system's perception depth and accuracy in drastically changing environments such as water and land crossings.

[0048] Optionally, the deep features include the estimated viscosity coefficient of the medium, the actuator load rate variance, and the sensor information entropy; the estimated viscosity coefficient of the medium is used to characterize the resistance characteristics of the current environment, the actuator load rate variance is used to quantify the degree of fluctuation of the actuator load, and the sensor information entropy is used to assess the uncertainty of the sensor data.

[0049] Specifically, to obtain an estimate of the medium's viscosity coefficient, the system utilizes a simplified robot dynamics model. This model correlates the robot's motion resistance in a specific medium with the medium's viscosity coefficient. The model is in the form of: ,in It is the resistance that the robot experiences. It is a drag coefficient related to the robot's shape. It is the viscosity coefficient of the medium to be estimated. The robot's movement speed is monitored in real time by a motor current sensor, which monitors the actuator's output torque. From this, the thrust applied to the medium is calculated, and this thrust is approximately equal to the resistance during stable movement. Meanwhile, the robot's actual speed is obtained through an inertial measurement unit or GPS. Due to the drag coefficient It is a constant calibrated experimentally beforehand, so the system can be solved at each time step. This allows for real-time updates of the estimated viscosity coefficient of the medium. Changes in this characteristic value directly reflect whether the robot is currently in air or water, or in a transitional region between the two, thus quantifying the drag characteristics of the current environment.

[0050] Furthermore, in order to calculate the actuator load rate variance, the system continuously collects current sensor data from each actuator (such as a motor) and converts it into a load rate percentage. The load rate is defined as the ratio of the current actual output torque to the rated maximum torque of the motor. Within a preset sliding time window, the system collects a series of load rate sampling values ​​and then calculates the statistical variance of these sampling values ​​within the window. The larger the variance, the more unstable the working state of the actuator and the more severe the load fluctuation. This usually corresponds to the robot experiencing a bumpy, turbulent, or unstable control process. This feature quantifies the degree of fluctuation in the actuator load and is an important intrinsic indicator for evaluating control smoothness and system stability.

[0051] To further quantify sensor information entropy, the system selects sensors crucial for robot navigation and control, such as gyroscopes or accelerometers in the inertial measurement unit. The system also collects the output signal sequence of these sensors within a sliding time window. To calculate the information entropy, the continuous sensor signals first need to be discretized, that is, the signal value range is divided into several non-overlapping intervals or "boxes." The frequency of the signal value falling into each interval within the time window is then counted, and this frequency is used as the probability of that interval occurring. The estimate, based on the definition of information entropy, is... Calculations show that a higher information entropy value means that the sensor signal is more evenly distributed and more unpredictable. This may indicate that the robot is in a state of violent movement or that the sensor is subjected to strong external noise interference. Sensor information entropy can quantify the uncertainty of sensor data.

[0052] By quantifying these three specific, deep-seated characteristics, the abstract concept of "overall system operational quality" is concretized and calculable. This enables subsequent decision-making systems to perceive environmental changes and assess their own state based on these precise, quantifiable indicators, rather than relying on fuzzy classifications or threshold judgments. These three characteristics, from the dimensions of external environmental resistance, internal actuator stability, and perceived information quality, together constitute a comprehensive and profound description of the dynamic characteristics of the robot-environment system, providing crucial, high-quality information input for achieving truly intelligent and adaptive control.

[0053] Optionally, the step of generating the dynamic value function includes: inputting the coupled feature vector into the neural network to obtain the structural parameters of the dynamic value function; and constructing the dynamic value function based on the structural parameters, wherein the dynamic value function includes penalty weights on the higher-order derivatives of the control quantity.

[0054] Furthermore, when constructing the dynamic value function, Lyapunov stability constraints are embedded to ensure that the dynamic value function can be guided to a stable controller.

[0055] Specifically, the coupled feature vector is used as input and passed to a pre-trained differentiable neural network. This network has been trained offline using a large amount of simulation or real-world data, learning the complex nonlinear mapping relationship between specific environmental and organismal state characteristics and the optimal control target weights. In real-time operation, the network can instantaneously output a set of structural parameters for a dynamic value function based on the input feature vector. These parameters are essentially a series of dynamically changing weight coefficients. The system uses these real-time generated structural parameters to construct the dynamic value function at the current moment. This function is typically designed as a weighted sum of multiple performance metrics, and its general structure can be expressed as follows:

[0056] ,

[0057] in, It is the output value of the dynamic value function; These are the structural parameter vectors output in real time by the neural network; they are dimensionless weight coefficients. It is the system's state vector, such as the robot's position, orientation, and velocity; It is the control input vector, such as the command force or torque of each actuator; This represents the tracking error term, which measures the deviation between the robot's current state and the desired state. This represents the energy consumption term, which is usually proportional to the square of the control input; This is a penalty item representing the agitation level that controls smoothness, among which... This is the derivative of the control input with respect to time, and this term is used to suppress drastic changes in the control quantity. For example, in chaotic regions transitioning between land and water, the coupled feature vectors reflect high uncertainty, causing the neural network output to have a large weight. This significantly increases the penalty for higher-order derivatives of the control quantity in the dynamic value function, prioritizing the stability of the motion.

[0058] To ensure the stability of the entire system under all operating conditions, Lyapunov stability constraints are imposed on the learning process of the pre-trained differentiable neural network during its offline training phase. This means that the optimization objective of the neural network is not only to match the data, but also to ensure that the dynamic value function constructed by any set of structural parameters of its output is consistent. Both can serve as a globally asymptotically stable Lyapunov function of the control system. This means that the function It must satisfy positive definiteness, and its time derivative must be... Under the influence of the control law, the value function must be negative definite. This constraint is achieved by adding a penalty term to the training loss function, thereby embedding stability into the value function generation mechanism from a theoretical level, ensuring that the adaptive adjustment target will never lead to system instability.

[0059] By utilizing neural networks to map low-level perceptual features to high-level control objectives, the core of the control strategy and the value function are dynamically constructed in real time. This solves the core problem of fixed control objectives and inability to adapt to drastic environmental changes in traditional methods. By embedding Lyapunov stability constraints, it provides a solid theoretical stability guarantee while endowing the system with a high degree of adaptability. This ensures that the robot will never sacrifice system stability while pursuing optimal energy efficiency or performance, thus forming the adaptive and safe cornerstone of the entire intelligent control method.

[0060] Optionally, the penalty weight for the higher-order derivative of the control quantity is automatically increased or decreased according to the chaotic or stable environment represented by the coupled eigenvector; the penalty weight is increased in a chaotic environment to suppress jitter, and the penalty weight is decreased in a stable environment to focus on energy consumption optimization.

[0061] Specifically, the core of this process lies in the internal mapping logic of the pre-trained differentiable neural network. During the design and training phases, this neural network learns the correlation between specific feature patterns in the coupled feature vectors and the priority of the optimal control objective. When the system detects that the robot is in a chaotic environment, such as in a water-land transition zone or encountering strong turbulence, the coupled feature vectors will exhibit a specific pattern, such as a sharp increase in both actuator load rate variance and sensor information entropy. This feature pattern is fed into the pre-trained differentiable neural network as input. The nonlinear activation function and weight matrix within the neural network map this input pattern into a specific set of outputs, namely the structural parameters of the dynamic value function. Under this specific input, the network is trained to significantly increase the weight coefficients corresponding to the higher-order derivative penalty terms of the control quantity, as described in the aforementioned formula. Weights in The dynamic value function automatically and significantly increases the penalty for drastic changes in control input. In the process of minimizing this value function, the inverter will be forced to generate an optimal control command that changes more smoothly and gently, thereby actively suppressing robot jitter and overshoot and placing system stability as the highest priority.

[0062] Conversely, when a robot enters a stable environment, such as cruising steadily on calm water or flat land, the coupling feature vector will exhibit low levels of actuator load variance and sensor information entropy. When this feature vector representing "stability" is input into a neural network, the network will output a different set of structural parameters accordingly. Specifically, the weights... The weights related to tracking error and energy consumption will be automatically reduced. and Then it will increase accordingly, such as Figure 2 As shown, this allows the optimization objective of the dynamic value function to automatically switch to focusing more on accurately tracking the predetermined trajectory and reducing energy consumption. The requirements for control smoothness are moderately relaxed, allowing the control system to take more direct and dynamic control behaviors to achieve optimization of energy efficiency and accuracy.

[0063] Through intelligent mapping of neural networks, a smooth switching between "stability priority" and "energy efficiency priority" modes for control objectives is achieved. This adaptive weight adjustment mechanism makes the dynamic value function no longer a static compromise target, but a "command stick" that can accurately reflect the core needs of the current task. This ensures that the robot can adopt the most appropriate control strategy in any environment. In dangerous environments, it automatically adopts conservative and robust behavior to ensure safety, while in safe environments, it automatically switches to an efficient and economical mode to improve work efficiency. Thus, dynamic synergistic optimization of stability and energy efficiency is achieved in dynamically changing environments.

[0064] Optionally, the steps to obtain the corrected dynamic value function include: inputting the dynamic value function and the coupled feature vector as initial conditions into the system evolution model to obtain time-varying constraints in the future time domain; deriving a correction amount for adjusting the structural parameters based on the time-varying constraints in the future time domain; and applying the correction amount to the structural parameters of the dynamic value function to obtain the corrected dynamic value function.

[0065] Specifically, based on the dynamic value function and coupled eigenvectors at the current moment, these are used as the initial conditions for the system evolution model. This system evolution model is an internally integrated dynamic model capable of simulating the future motion trends of the amphibious robot. Based on the current control tendency (implied by the dynamic value function) and environmental characteristics (represented by eigenvectors), the model forward extrapolates the robot's state over a short future time domain, thereby predicting time-varying constraints in the future time domain. These time-varying constraints are the physical or functional limitations the robot will face. For example, when the model predicts that the required torque of a certain actuator will exceed its maximum rated value, an "actuator saturation boundary" constraint will be triggered. Based on the predicted future time-varying constraints, the system uses a reverse correction mechanism to generate a correction amount to adjust the current objective. The core of this reverse process is establishing a sensitivity relationship between the probability of future constraints being triggered and the structural parameters of the current dynamic value function. For example, if the system analyzes that the risk of future actuator saturation mainly stems from the excessive weighting of the energy consumption term in the current dynamic value function, leading to an overly aggressive control strategy, the system will generate a correction amount. This correction amount is a vector with the same dimension as the structural parameter vector, with a negative value for the weighted component corresponding to the energy consumption term and a positive value for the weighted component corresponding to the control smoothness term. The formula for generating this correction amount can be expressed as follows:

[0066] ,

[0067] in, It is the generated correction vector; It is a positive learning rate; It is a cost function that quantifies the degree to which future constraints are violated, such as the amount by which an actuator exceeds its saturation boundary; This represents the structural parameters of the cost function relative to the current dynamic value function. The gradient of this gradient indicates the direction for adjusting the current weights to reduce future risk.

[0068] The system uses this correction amount to adjust the original structural parameters of the dynamic value function generated in the previous stage, obtaining the corrected dynamic value function. This adjustment process is usually a simple vector addition operation, i.e., the corrected structural parameters... equal to the original parameter Add correction amount Therefore, the corrected dynamic value function, although still used in the current moment, implicitly incorporates strategies to avoid future constraints in its internal weight allocation. For example, because the weight of smoothness is increased in advance, the robot's control behavior becomes smoother and more stable before it is fully submerged in water, preparing for the upcoming loss of GPS signal in terms of posture and sensory dependence.

[0069] By introducing a closed loop for predicting and correcting future time-varying constraints, the "short-sightedness" problem prevalent in traditional adaptive control is successfully solved. It extends the robot's decision-making from the current moment to the future time domain, avoiding the risk of falling into future dilemmas while pursuing local optima. This forward-looking correction mechanism makes the dynamic value function more intelligent and holistic in guiding robot behavior, significantly improving the smoothness, safety, and overall energy efficiency of the robot during complex task transitions such as crossing land and water domains.

[0070] Optionally, the time-varying constraints in the future time domain include actuator saturation boundary, sensor availability, and energy recovery potential; the actuator saturation boundary is used to warn of the upper limit of control capability, sensor availability is used to warn of the loss of sensing information, and energy recovery potential is used to identify the external energy that can be used in the future.

[0071] Specifically, for predicting actuator saturation boundaries, the system evolution model extrapolates the theoretical control output required by each actuator (such as propulsion motors, servo motors, etc.) in the short-term future based on the current state and control trends. The system internally stores the physical performance limits of each actuator, namely its maximum torque, speed, or power. During the prediction process, the system continuously compares the predicted theoretical requirements with the physical limits. Once the model predicts that at some point in the future, the theoretical requirements of an actuator will exceed its maximum capacity, the system will mark this event as an impending "actuator saturation boundary" constraint. Furthermore, for predicting sensor availability, the system evolution model is combined with an environmental sensor failure model. This model stores the working states of different sensors under specific environmental conditions. For example, if the robot is equipped with a vision sensor, the model may include a rule that triggers "vision sensor failure" when the turbidity of the water exceeds a certain threshold. This prediction allows the system to know in advance about the impending loss of perception capabilities.

[0072] Meanwhile, for predicting energy recovery potential, the system evolution model will combine external environmental data (such as water flow information obtained through flow meters or wave data predicted by wave models) for analysis. For example, when the model predicts that the robot will enter an area with downstream navigation in the future, it will identify this as an opportunity to save energy by appropriately reducing the active propulsion force, or even to recover energy by utilizing the water flow. Under certain wave conditions, the robot can obtain energy from the undulations of the waves through specific attitude control. When the model identifies such conditions, it will generate a prediction of "energy recovery potential". This is essentially a favorable constraint, prompting the control system to adjust its strategy to utilize external energy.

[0073] By accurately predicting these three types of specific time-varying constraints, the system's forward-looking capabilities are enriched. This enables the system to not only combat future negative risks (such as saturation and failure), but also proactively identify and utilize future positive opportunities (such as energy recovery). This comprehensive and multi-dimensional prediction of future constraints provides detailed and specific information input for subsequent dynamic value function correction, ensuring that correction decisions are targeted and refined. Thus, in complex and ever-changing cross-domain operating environments, energy efficiency and performance optimization are elevated to a new level of strategic planning.

[0074] Optionally, the steps for solving the optimal control command include: using the corrected dynamic value function as the objective function to construct a nonlinear model predictive control optimization problem; performing sensitivity analysis based on the optimal solution at the previous time step to generate the initial iteration point of the current optimization problem; and iteratively solving from the initial iteration point to obtain the optimal control command.

[0075] Specifically, at the beginning of each control cycle, the system uses the corrected dynamic value function as the core objective function and inputs it into a real-time nonlinear model predictive control solver. Within each control cycle, this solver solves an optimization problem aimed at minimizing the cumulative value over a finite future prediction time domain. The objective function of this optimization problem is... It can be represented as

[0076] ,

[0077] The summation range is from the current time. To the end of the prediction time domain; It is the corrected dynamic value function; Is this function in Structural parameters at time; and These represent the predicted system state and control input at time t, respectively. Considering that the structural parameters of the dynamic value function may change drastically in each control cycle, causing the optimization problem itself to change rapidly, traditional optimization methods starting from scratch are difficult to meet real-time requirements. Therefore, this method adopts a hot-start strategy based on the sensitivity analysis of the optimal solution at the previous time step. By analyzing the differential relationship between the value function at the previous and current time steps, the approximate location of the optimal solution for the current optimization problem is predicted, thereby quickly generating a high-quality initial iteration point for the current optimization problem. This reduces the number of iterations required for the optimization algorithm to search from a random or fixed initial point. Based on this initial iteration point, the optimization solver efficiently converges and solves a set of optimal control sequences in the future prediction time domain through iterative calculations, such as using sequential quadratic programming or interior-point methods. The system finally extracts the first instruction from this sequence as the optimal control instruction for the current time step and sends it to the robot's actuators, such as motors or thrusters, to complete the actual physical action. This process is repeated in each control cycle, forming a continuous rolling optimization.

[0078] As the final execution stage of the entire control method, its key role is to accurately and efficiently transform the highly abstract and dynamically changing control intentions generated in the first two steps into concrete control actions in the physical world. By introducing a hot-start strategy specifically designed for time-varying objective functions and based on sensitivity analysis, it successfully overcomes the enormous computational challenges of solving complex nonlinear optimization problems online in real time, significantly improving the real-time response capability of the control system. This ensures that the robot can closely follow the constantly evolving "optimal" behavior criterion defined by the dynamic value function, such as... Figure 3 As shown, this enables the adaptive capabilities and forward-looking planning of the entire system to be realized, forming an indispensable decision-making and execution link in the entire closed-loop control system.

[0079] Optionally, the steps of constructing a nonlinear model predictive control optimization problem include: dynamically constructing a set of constraints for the real-time optimization problem based on the structural parameters of the coupled eigenvectors and the corrected dynamic value function; when the control smoothness penalty weight in the structural parameters is greater than a first threshold, tightening the actuator peak torque constraint from the rated torque value to a lower robust torque value; and when the estimated value of the medium viscosity coefficient in the coupled eigenvectors indicates a medium switching, adding an idling constraint to limit the thruster idling speed to the set of constraints.

[0080] Specifically, this process is based on the judgment and adjustment of the structure parameters of the coupled feature vector and the corrected dynamic value function. The specific operation includes two parallel adaptive constraint mechanisms. The first mechanism is the dynamic tightening of the actuator peak torque. The system monitors in real time the magnitude of the weight representing the control smoothness penalty in the structure parameters of the corrected dynamic value function. When this weight value, such as its matrix norm, exceeds a preset first threshold, the system determines that it is currently in a chaotic environment where jitter needs to be suppressed. The actuator peak torque constraint in the nonlinear model predictive control optimization problem will automatically and proactively tighten from the rated torque value specified in its hardware specifications to a lower preset threshold. The first mechanism involves setting a robust torque value to physically limit the optimizer from generating overly aggressive control commands, ensuring smooth operation in unstable environments. The second mechanism addresses specific risk management for cross-medium operations. The system continuously monitors the estimated viscosity coefficient of the medium in the coupled eigenvectors. When the rate of change or absolute value of this estimate abruptly exceeds a certain threshold, it indicates that the robot is experiencing or about to experience a medium switch, such as emerging from water. In this case, the thruster faces the risk of a sudden shift from high to low load. To prevent the thruster from spinning at ultra-high speeds in air due to loss of resistance, a specific idling constraint is dynamically added to the constraint set of the optimization problem. This constraint sets a temporary upper limit on the thruster's rotational speed, significantly lower than its limiting speed, thereby limiting the thruster's idling speed at the moment of medium switching.

[0081] By transforming the robot's operational constraints from static boundaries to dynamic, environment-adaptive intelligent safety barriers, this dynamic constraint construction capability significantly enhances the robot's safety and robustness. In harsh environments, by proactively tightening torque constraints, the system can preventively avoid violent movements that could lead to instability, demonstrating a conservative strategy based on risk assessment. At critical nodes crossing land and water, by precisely adding idling constraints, the system can calmly cope with the physical challenges brought about by sudden changes in the medium, avoiding drastic energy waste and potential damage to actuator hardware. This forward-looking constraint management, deeply coupled with operating conditions, ensures that energy efficiency optimization is always carried out within dynamically adjusted safety boundaries, improving the robot's overall survivability and operational reliability in complex and ever-changing task scenarios.

[0082] Optionally, the steps of updating the model used for feature extraction and prediction online include: comparing the actual system response data with the predicted response when the optimal control command is issued to generate a prediction error; and using the prediction error to fine-tune the feature extraction model used to generate coupled feature vectors and the system evolution model for predicting time-varying constraints online.

[0083] Specifically, the system continuously optimizes its internal model by utilizing real-world feedback. This process begins with the actual execution of the optimal control commands generated in the previous stage and sent to the actuators. Simultaneously, the system collects a series of real-world response data reflecting the robot's actual performance. This data is multi-dimensional, including but not limited to the actual energy consumption recorded by the battery management system, the actual motion trajectory and attitude changes recorded by the inertial measurement unit and GPS, and the actual signal quality and noise levels from various sensors.

[0084] After data acquisition, the system feeds back the actual system response data in two ways. The first way is fed back to the feature extraction model responsible for generating the coupled feature vector. The system compares the actual response data with the response predicted by the internal model when the optimal control command is issued, forming a prediction error signal. For example, if the actual energy consumption is significantly higher than the energy consumption predicted based on the coupled feature vector at that time, the error will be used to fine-tune the parameters in the feature extraction model online. This fine-tuning process can use adaptive algorithms such as recursive least squares. The purpose is to adjust the parameters inside the model so that the coupled feature vector generated later can more accurately reflect the real dynamic interaction between the system and the environment, such as more accurately estimating the viscosity coefficient of the medium.

[0085] The second feedback loop sends the actual system response data to the system evolution model responsible for predicting future time-varying constraints. The system compares the actual constraint event, such as the loss of GPS signal at a certain depth, with the depth at which the model initially predicted the event would occur. The deviation between the prediction and reality is used as a correction signal to update the internal parameters of the system evolution model. In this way, the prediction model can continuously learn and adapt to the real characteristics of the environment, such as the actual attenuation of signals by water, thereby making its ability to predict future time-varying constraints more accurate and reliable over time.

[0086] By establishing a feedback loop from physical execution to model updates, the entire control system is endowed with the ability to learn and self-optimize online. This ensures that the system's internal world model does not become detached from physical reality due to initial errors or gradual environmental changes. This continuous self-correction mechanism enhances the system's robustness to model uncertainties and external unknown disturbances, enabling the amphibious robot not only to passively adapt to the environment but also to actively learn from it. As operating time increases, its energy efficiency optimization and decision-making level will continuously improve, ultimately achieving true long-term autonomous and intelligent control.

[0087] In this embodiment, to verify the feasibility of the invention in practice, it is applied to an amphibious cross-domain operation robot named "AR-200". The robot's typical task is to perform reciprocating exploration between a lake and the surrounding land. This task requires the robot to traverse areas with drastically changing physical characteristics, including shorelines, shallow water, and deep water. Traditional control methods typically employ segmented controllers based on fixed parameters, which often fail to adapt to sudden environmental changes in water-land transition zones, leading to motion instability, increased energy consumption, and even mission failure. This embodiment aims to verify the superiority of the method of this invention compared to traditional methods.

[0088] In this embodiment, the invention first acquires and processes multimodal sensor data. When the robot moves from land to water, its onboard motor current sensor, inertial measurement unit (IMU), GPS, and water depth sensor collect data in real time. The system performs coupled analysis on this data: by combining the motor output torque and the speed measured by the IMU, the viscosity coefficient of the medium is estimated in real time. When moving on land, this estimated value is close to the air viscosity coefficient; when the robot's wheels contact the water surface and gradually submerge, this value increases smoothly and rapidly, accurately characterizing the change in environmental resistance. Simultaneously, within a 1-second sliding time window, the system calculates the variance of the motor load rate and the information entropy of the IMU angular velocity signal. On bumpy shores and when splashing occurs upon entering the water, both the actuator load rate variance and the sensor information entropy increase significantly, reflecting the high degree of uncertainty and chaos in the system state. These three deep features—the estimated viscosity coefficient of the medium, the actuator load rate variance, and the sensor information entropy—are combined into a low-dimensional coupled feature vector.

[0089] Next, the feature vector is input into a pre-trained differentiable neural network with embedded Lyapunov stability constraints to generate a dynamic value function. Experimental data show that when the feature vector reflects shore turbulence and water inrush (both actuator load variance and sensor information entropy are high), the neural network automatically outputs a set of weights that increase the control smoothness penalty. Structural parameters. For example, in the transition phase. The weight value was dynamically increased from 0.2 when the robot was stable on land to 1.5. This automatically switched the optimization objective of the value function to "stability priority," making the controller's output commands smoother and suppressing the robot's violent posture jitters upon entering the water. Once the robot was fully in calm water, the variance and entropy values ​​in the feature vector decreased, and the neural network then... The weight is reduced to 0.3, while the tracking error weight is increased. and energy consumption weight This shifts the control objective to "prioritizing energy efficiency and precision".

[0090] As the robot approaches the water's edge, the forward-looking correction mechanism of this invention is activated. Based on the current state and the value function tendency, the system evolution model predicts that the robot will trigger a sensor availability constraint of "GPS signal loss" within the next 5 seconds due to the antenna submerging in the water. To cope with this foreseeable loss of sensory information, the system generates a correction value in reverse. This correction amount is used to moderately increase the smoothness weight in advance. This allowed the robot to enter an extremely stable navigation attitude before the GPS signal was completely lost, preparing it for subsequent inertial navigation relying on the IMU. Simultaneously, the model also predicted that if it entered the water at the current speed, the torque demand of the subsequent thrusters would exceed the rated value by 120%, triggering the "actuator saturation boundary" constraint. Based on this, the correction amount... It also includes weighting for energy consumption. The increase in size allows for a gradual reduction in the robot's entry speed into the water, thus preventing overload of the actuator.

[0091] The dynamic value function, after look-ahead correction, is used as the target and optimized in real time by a nonlinear model predictive control (NMPC) solver. Thanks to the hot-start strategy based on sensitivity analysis of the solution from the previous time step, even during the water entry phase when the structural parameters of the value function change drastically, the single-step solution time of the NMPC solver can still be kept within 20 milliseconds, which meets the requirements of real-time control. The first instruction of the optimal control sequence is sent to the robot's wheel motors and underwater thrusters.

[0092] Finally, the closed-loop learning mechanism continues to play a role in the task execution process. In the first few cross-domain tasks, the system found that the actual energy consumption was 8% higher than the value predicted based on the initial model. These prediction errors were fed back to update the fluid dynamics parameters in the feature extraction model and the sensor failure model parameters in the system evolution model online. After about 10 task cycles, the prediction error of the model was significantly reduced, which further improved the energy efficiency and transition smoothness of subsequent tasks.

[0093] Comparing the average data from 50 repeated experiments, this invention demonstrates significant technical effectiveness. The peak attitude angular velocity of the experimental group robot during the water-land transition phase was reduced by approximately 40% compared to the control group, and the actuator load rate variance was reduced by more than 60%, thus avoiding the risk of instability. Figure 4 As shown, due to forward-looking speed adjustments and dynamic energy efficiency optimization, the experimental group reduced the average total energy consumption for completing a single cross-domain task by approximately 18% compared to the control group.

[0094] It should be noted that the electrical connections between the various units described above do not necessarily represent direct or indirect connections. Any indirect connection method can be applied to the embodiments of the present invention as long as it achieves the purpose of the present invention. The above are merely exemplary embodiments of the present invention and should not be construed as limiting the scope of the present invention.

[0095] All equivalent changes and modifications made in accordance with the teachings of this invention are still within the scope of this invention. Those skilled in the art will readily conceive of other embodiments of this invention upon considering the specification and the disclosure of practical truth. This application is intended to cover any variations, uses, or adaptations of this invention that follow the general principles of this invention and include common knowledge or conventional techniques in the art not described herein.

Claims

1. An amphibious robot energy efficiency optimization control method for water-land cross-domain operation, characterized in that, The method comprises: acquiring the multi-modal sensor data of the amphibious robot, and performing feature extraction on the multi-modal sensor data to generate a coupling feature vector; based on the coupling feature vector, generating a dynamic value function by using a neural network trained with stability constraints, the structural parameters of the dynamic value function automatically adjusting with changes in the environment; based on the dynamic value function and the coupling feature vector, predicting time-varying constraints in a future time domain by a system evolution model, and generating a correction amount to adjust the structural parameters of the dynamic value function to obtain a corrected dynamic value function; performing real-time nonlinear optimization on the corrected dynamic value function to solve optimal control instructions; executing the optimal control instructions and collecting system actual response data, and using the system actual response data to update the models used for feature extraction and prediction online.

2. The method of claim 1, wherein, The step of performing feature extraction on the multi-modal sensor data to generate a coupling feature vector comprises: separating internal state data and external environment data from the multi-modal sensor data; performing coupling analysis on the internal state data and the external environment data to extract deep features reflecting the overall operation quality of the system; and generating the coupling feature vector based on the deep features.

3. The method of claim 1, wherein, The step of generating a dynamic value function comprises: inputting the coupling feature vector into the neural network to obtain the structural parameters of the dynamic value function; and constructing the dynamic value function according to the structural parameters, the dynamic value function including a penalty weight for the high-order derivative of the control amount.

4. The method of claim 1, wherein, The step of obtaining a corrected dynamic value function comprises: inputting the dynamic value function and the coupling feature vector as initial conditions into the system evolution model to obtain time-varying constraints in the future time domain; reversely deriving the correction amount for adjusting the structural parameters according to the time-varying constraints in the future time domain; and applying the correction amount to the structural parameters of the dynamic value function to obtain the corrected dynamic value function.

5. The method of claim 1, wherein, The step of solving optimal control instructions comprises: constructing a nonlinear model predictive control optimization problem by taking the corrected dynamic value function as the objective function; performing sensitivity analysis based on the optimal solution at the last time to generate an initial iteration point of the current optimization problem; and starting iteration from the initial iteration point to obtain the optimal control instructions.

6. The method of claim 1, wherein, The step of updating the models used for feature extraction and prediction online comprises: comparing the system actual response data with the predicted response when the optimal control instructions are issued to generate a prediction error; and using the prediction error to fine-tune the model used for generating the coupling feature vector and the system evolution model used for predicting the time-varying constraints online.

7. The method of claim 2, wherein, The deep features include a medium viscosity coefficient estimate, an actuator load rate variance, and a sensor information entropy; the medium viscosity coefficient estimate is used to represent the resistance characteristics of the current environment, the actuator load rate variance is used to quantify the fluctuation degree of the actuator load, and the sensor information entropy is used to evaluate the uncertainty of the sensor data.

8. The method of claim 3, wherein, The penalty weight of the high order derivative of the control variable is automatically increased or decreased according to the chaotic or stable environment represented by the coupling eigenvector; the penalty weight is increased in the chaotic environment to suppress chattering, and the penalty weight is decreased in the stable environment to focus on energy consumption optimization.

9. The method of claim 4, wherein, The time-varying constraints in the future time domain include actuator saturation boundaries, sensor availability and energy recovery potential; the actuator saturation boundaries are used to warn the upper limit of control capability, the sensor availability is used to warn the loss of perception information, and the energy recovery potential is used to identify the external energy available in the future.

10. The method of claim 5, wherein, The step of constructing a nonlinear model predictive control optimization problem includes dynamically constructing a constraint condition set of a real-time optimization problem based on the coupling eigenvector and the structural parameters of the corrected dynamic value function; when the control smoothness penalty weight in the structural parameters is greater than a first threshold value, the actuator peak torque constraint is tightened from the rated torque value to a lower robust torque value; When the medium viscosity coefficient estimate value in the coupling eigenvector indicates medium switching, an idling constraint for limiting the idling speed of the propeller is added to the constraint condition set.

Images