Real-time cooperative control method and system for robot parkour action

Abstract

The application relates to the technical field of robot motion control, and discloses a real-time cooperative control method and system for robot parkour motion, wherein the method comprises the following steps: acquiring a trunk angular velocity and a joint angular velocity in an air flight stage, calculating an angular momentum residual error vector and inputting the vector into a causal convolution time series prediction network to generate a landing posture deviation estimation vector; calculating a double-leg asymmetric impact force pre-distribution vector based on the vector; pre-calculating a joint impact wave propagation prediction parameter table; generating a double-leg differential cascading buffer pre-scheduling scheme by a dynamic programming algorithm and preloading the scheme to joint impedance controllers; and performing online correction on upstream joint parameters according to real-time torque feedback after landing.

Description

Technical Field

[0001] This invention relates to the field of robot motion control technology, specifically to a real-time collaborative control method and system for robot parkour movements. Background Technology

[0002] In the field of motion control for robot parkour, the flight phase and the landing cushioning phase are typically handled by separate control modules. The flight phase adjusts the torso posture through limb swings, while the landing cushioning phase calculates impedance buffering parameters by detecting the propagation delay of the impact force signal along the joint chain.

[0003] In the existing scheme, there is a lack of real-time information coordination between the two-stage control modules, and the attitude deviation prediction information acquired in the air phase is not fed forward to the landing buffer scheduling stage. When the air phase predicts a roll angle deviation in the torso upon landing, the actual landing impact exhibits an asymmetrical force amplitude distribution between the left and right legs, and the impact propagation path characteristics of the tilted leg differ from those of a normal vertical landing. The landing buffer control module only initiates propagation delay detection and scheduling calculations after the impact occurs. During the time waiting for the detection to complete, the ankle joint has already endured excessive load, and the buffering strategy cannot match the asymmetrical impact distribution, leading to an imbalance in energy distribution between end-joint overload and insufficient proximal joint buffering. Summary of the Invention

[0004] This invention provides a real-time collaborative control method and system for robot parkour movements, solving the technical problems in related technologies where robots suffer from poor landing stability and high risk of joint overload due to unpredictable posture deviations, unreasonable distribution of impact force between legs, and lack of coordination in the timing of buffering of various joints when landing during parkour.

[0005] This invention discloses a real-time collaborative control method for robot parkour movements, including: during the flight phase, calculating the angular momentum residual vector based on the torso angular velocity and joint angular velocity, inputting the angular momentum residual vector sequence into a causal convolutional temporal prediction network, and outputting a landing attitude deviation prediction vector. Based on the roll angle deviation and pitch angle deviation in the landing posture deviation prediction vector, the asymmetric impact force pre-distribution vector of the two legs is calculated in combination with the robot kinematics model. For the expected impact force amplitude of each leg in the pre-allocation vector of the asymmetric impact force of the two legs, based on the structural equivalent longitudinal wave propagation velocity parameter and the equivalent impedance attenuation coefficient, an independent shock wave propagation prediction parameter table for the left and right legs is generated. Based on the shock wave propagation prediction parameter table, a dynamic programming algorithm is used to calculate the buffer start-up time offset and time-varying impedance parameter curve for each joint, generating a differentiated cascaded buffer pre-scheduling scheme for the two legs. The pre-scheduling scheme is preloaded into the impedance controller of each joint. Upon landing, each joint enters a compliant buffer state according to the pre-configured parameters, and online correction is performed on the upstream joints that have not yet started buffering based on real-time torque feedback.

[0006] Furthermore, the calculation method for the angular momentum residual vector includes: In each control cycle, based on the angular velocities of each joint and the trunk, the angular velocity contributions of each body segment are summed using the composite inertia matrix of the whole-body rigid tree structure to obtain the three-dimensional vector of the measured angular momentum of the system. The three-dimensional vector of the measured angular momentum of the system is then subtracted component by component from the reference value of angular momentum recorded at the take-off moment to obtain the residual vector of angular momentum for the current cycle.

[0007] Furthermore, the causal convolutional temporal prediction network is a one-dimensional temporal convolutional network, whose convolutional kernels only apply operations to data at the current time and historical time. The input layer of the causal convolutional temporal prediction network receives a sequence of angular momentum residual vectors from the start time to the current control cycle. The input at each time step in the sequence is a three-dimensional angular momentum residual vector. The output layer is a fully connected layer that outputs a four-dimensional bias vector, which includes roll bias, pitch bias, roll velocity bias and pitch velocity bias. Before inputting the causal convolutional temporal prediction network, each component of the angular momentum residual vector sequence is normalized using Z-score, and the network output is denormalized to restore the physical dimensions.

[0008] Furthermore, the pre-distribution vector of asymmetric impact force for both legs includes the expected contact time difference and the expected normal impact force amplitude for each leg; wherein, the calculation method for the expected contact time difference includes: Based on roll angle deviation and hip width Calculate the height difference between the left and right feet. Combined with the current vertical falling speed Calculate the time difference between the contact time of the left and right feet. .

[0009] Furthermore, the distribution method of the expected normal impact force amplitude of each leg includes: The offset of the centroid projection in the left and right support regions is determined based on the pitch angle deviation and roll angle deviation. The proportion of the normal impact force borne by the inclined leg is calculated according to the moment balance relationship. The proportion of normal impact force borne by the opposite leg is ,in The lateral distance between the two feet. Due to roll angle deviation and pitch angle deviation The predicted normal impact force amplitude of each leg is calculated using the center-of-mass position equation in the robot's kinematic model; the predicted normal impact force amplitude of each leg is equal to the robot's total weight. Multiply by the corresponding proportion.

[0010] Furthermore, the shock wave propagation prediction parameter table contains three sets of parameter records for each leg, corresponding to the ankle, knee, and hip joints respectively. Each set of parameter records includes the expected time delay value and the expected peak torque amplitude value; wherein: The expected time delay of each joint is equal to the equivalent structural path length from the foot to that joint divided by the equivalent longitudinal wave propagation velocity parameter of the structure; The expected peak torque amplitude of each joint is ,in The expected impact force amplitude at the foot of the leg. For the first The equivalent impedance attenuation coefficient of the segment connection structure at the current joint angle. Take 1, 2, and 3 in sequence, corresponding to the ankle joint, knee joint, and hip joint, respectively.

[0011] Furthermore, the dynamic programming algorithm aims to uniformly distribute the actual impact energy absorbed by each joint across all joints in the body, with the objective function being: ; in Number the joints. For joints The actual energy absorbed during the buffering process The total impact energy absorbed by the six joints; The constraints include: ,in For joints The expected peak torque amplitude, For joints The torque limit value; ,in For the joint during the cushioning process angular change This is the upper limit of the angle travel; ,in The angular velocity at which the torso rolls back to the neutral position after landing. This represents the lower limit of the recovery rate.

[0012] Furthermore, the time-varying impedance parameter curve includes the stiffness and damping values ​​of each joint as a function of time during the buffering process.

[0013] The offset of the cushioning initiation moment of the tilted leg ankle joint is negative, indicating that it enters a compliant state earlier than the actual foot contact moment; the initial damping value of the tilted leg ankle joint is higher than the initial damping value of the contralateral leg ankle joint.

[0014] Furthermore, the online correction includes: At each joint where the buffer has been activated, the current torque feedback value is obtained, and the difference between the current torque feedback value and the expected torque value at the corresponding moment in the pre-scheduling scheme is calculated as the torque deviation. When the absolute value of the torque deviation exceeds a preset threshold, the torque deviation value is transmitted to the time-varying impedance parameter curve corresponding to the upstream joint that has not yet started the buffer; if the torque deviation is positive, the initial damping value of the upstream joint is increased and its buffer start time is advanced.

[0015] If the torque deviation is negative, the initial damping value of the upstream joint is reduced; the corrected parameters are written into the parameter buffer of the corresponding joint impedance controller. The parameter buffer is a double-buffered structure that supports receiving updated parameters while the current parameters are being executed.

[0016] This invention provides a real-time collaborative control system for robot parkour movements, comprising: The attitude deviation prediction module is used to calculate the angular momentum residual vector based on the torso angular velocity and joint angular velocity during the flight phase. The angular momentum residual vector sequence is input into the causal convolutional temporal prediction network to output the landing attitude deviation prediction vector. The impact force pre-distribution module is used to calculate the asymmetric impact force pre-distribution vector of the two legs based on the roll angle deviation and pitch angle deviation in the landing posture deviation prediction vector, combined with the robot kinematic model. The propagation prediction module is used to generate independent shock wave propagation prediction parameter tables for the left and right legs based on the structural equivalent longitudinal wave propagation velocity parameters and the equivalent impedance attenuation coefficient for the expected impact force amplitude of each leg. The buffer scheduling module is used to calculate the buffer start-up time offset and time-varying impedance parameter curve for each joint based on the shock wave propagation prediction parameter table and through dynamic programming algorithm, and generate a differentiated cascaded buffer pre-scheduling scheme for the two legs.

[0017] The execution correction module is used to preload the pre-scheduling scheme into the impedance controller of each joint, control each joint to enter a compliant buffer state according to the pre-configured parameters when landing, and perform online correction on the upstream joints that have not yet started buffering based on real-time torque feedback.

[0018] Compared with the prior art, the present invention has the following beneficial effects: This invention establishes a feedforward transmission channel for the landing attitude deviation prediction vector between the airborne flight phase and the landing buffer phase, enabling the differentiated buffering parameters of each joint to be pre-configured and loaded onto the corresponding impedance controller before landing. The tilted ankle joint receives an earlier buffering activation time and a higher initial damping value before ground contact, ensuring the impedance controller is already in a compliant state upon impact. This eliminates the scheduling lag caused by initiating propagation delay detection only after landing, preventing the ankle joint from bearing excessive load during the delay detection period. Independent shock wave propagation prediction parameter tables for the left and right legs match the buffering strategy of each joint with the impact force amplitude it bears, improving the energy distribution imbalance problem of overload at the distal joint and insufficient buffering at the proximal joint under a unified buffering strategy. Real-time torque deviation correction after landing allows the pre-scheduling scheme to maintain effective buffering through online compensation of upstream joints even when residual errors exist in the prediction, improving the robustness of the overall buffering control. Attached Figure Description

[0019] Figure 1 This is a flowchart of a real-time collaborative control method for robot parkour movements provided in an embodiment of the present invention;

[0020] Figure 2 This is a schematic diagram of the cumulative trend of the angular momentum residual vector during the flight phase provided in an embodiment of the present invention;

[0021] Figure 3 This is a schematic diagram of the landing attitude deviation prediction vector (output of the 18th control cycle) provided in an embodiment of the present invention;

[0022] Figure 4 This is a schematic diagram comparing the asymmetrical distribution ratio of impact force between the two legs according to an embodiment of the present invention;

[0023] Figure 5 This is a schematic diagram comparing the peak amplitude of the expected torque of each joint of the left and right legs according to an embodiment of the present invention;

[0024] Figure 6 This is a schematic diagram of the initial damping value distribution of the pre-scheduled dual-leg differentiated cascaded buffer provided in an embodiment of the present invention;

[0025] Figure 7 This is a schematic diagram of the distribution of the initial stiffness value of the pre-scheduled dual-leg differentiated cascaded buffer provided in an embodiment of the present invention;

[0026] Figure 8 This is a schematic diagram comparing the knee joint damping values ​​before and after online correction, provided in an embodiment of the present invention.

[0027] Figure 9 This is a schematic diagram comparing the expected and measured peak torque of the ankle joint according to an embodiment of the present invention. Detailed Implementation

[0028] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0029] Example

[0030] A real-time collaborative control method for robot parkour movements according to an embodiment of the present invention, such as... Figure 1 As shown, it includes the following steps: Step 1: Obtain the angular momentum residual during the flight phase and predict the landing attitude deviation; During each control cycle of the flight phase, the torso angular velocity output by the inertial measurement unit and the joint angular velocities output by the encoders of each joint are acquired. Based on the whole-body inertia matrix, the measured angular momentum vector of the system is calculated and subtracted from the reference angular momentum value recorded at takeoff to generate an angular momentum residual vector. The accumulated angular momentum residual vector time series is input into a causal convolutional temporal prediction network, which outputs the predicted torso roll angle deviation, pitch angle deviation, and corresponding angular velocity deviation at the landing moment, generating a landing attitude deviation prediction vector.

[0031] It should be noted that the above-mentioned angular momentum residual vector is calculated as follows: In each control cycle, based on the angular velocities of each joint and the trunk, the angular velocity contribution of each body segment is summed using the composite inertia matrix of the whole-body rigid tree structure to obtain the system's measured three-dimensional angular momentum vector; the system's measured three-dimensional angular momentum vector is then subtracted component by component from the angular momentum reference value frozen and recorded at the takeoff moment to obtain the angular momentum residual vector for the current cycle.

[0032] It should be noted that the aforementioned causal convolutional temporal prediction network is a one-dimensional temporal convolutional network. Its convolutional kernels only apply operations to data from the current and historical moments, without involving data from future moments. The input layer of the causal convolutional temporal prediction network receives a sequence of angular momentum residual vectors from the take-off moment to the current control cycle. The input at each time step in the sequence is a three-dimensional angular momentum residual vector. The output layer is a fully connected layer, outputting a four-dimensional bias vector, including roll angle bias, pitch angle bias, roll rate bias, and pitch rate bias, i.e., the landing attitude bias prediction vector. During training, the entire process data from parkour take-off to landing collected in the simulation environment is used as the training set. The measured torso roll angle bias, pitch angle bias, and corresponding angular rate bias at the landing moment are used as supervision labels. The mean squared error loss function is used, and the Adam optimization algorithm is used for parameter updates. In the landing attitude deviation prediction vector, the units of roll angle deviation and pitch angle deviation are radians, and the units of roll angular velocity deviation and pitch angular velocity deviation are radians per second. Before inputting into the causal convolutional temporal prediction network, each component of the angular momentum residual vector sequence is Z-score normalized to eliminate the influence of differences in the magnitude of different components on network training. After denormalization, each component of the landing attitude deviation prediction vector output by the causal convolutional temporal prediction network is restored to its physical dimensions for use in subsequent steps.

[0033] Step 2: Calculate the pre-assigned vector of asymmetric impact force on both legs; Based on the roll and pitch deviations in the landing attitude deviation prediction vector, the differences in the expected contact time of each foot relative to the ground and the expected normal impact force distribution ratio are calculated using the robot's kinematic model. The tilted foot is expected to contact the ground earlier than the opposite foot and bear a larger proportion of the normal impact force. An asymmetric impact force pre-distribution vector for both legs is generated.

[0034] It should be noted that the above-mentioned calculation method for the difference in estimated contact time is as follows: based on the roll angle deviation and the current hip joint width, the torso roll angle deviation is projected as the height difference between the left and right feet; combined with the current estimated descent velocity, the height difference is converted into the difference in the timing of the left and right feet contacting the ground. Specifically, let the roll angle deviation be... (Unit: radians), hip joint width is (Unit: meters), the current estimated falling speed is... (Unit: meters per second, absolute value), then the height difference between the left and right feet is... (Unit: meters), the time difference between contact between the left and right feet is... (Unit: seconds).

[0035] It should be noted that the above-mentioned calculation method for the expected normal impact force distribution ratio is as follows: the offset of the center of mass projection in the left and right support regions is determined based on the pitch angle deviation and roll angle deviation. (Unit: meters), the proportion of the normal impact force borne by each leg is distributed according to the torque balance relationship. Specifically, let the total weight of the robot be... (Unit: Newtons), the lateral distance between the two feet is... (Unit: meters), then the proportion of the normal impact force borne by the inclined leg is: The proportion of normal impact force borne by the opposite leg is Both ratios are dimensionless, where the offset is... Due to roll angle deviation and pitch angle deviation It is obtained by calculating the position equation of the center of mass in the robot's kinematic model. For pitch angle deviation, , These are inherent parameters of the robot body. The predicted normal impact force amplitudes (i.e., the values ​​of the tilted leg and the opposite leg) are also considered. Multiplied by the corresponding ratio (unit: Newtons) and the difference from the expected contact time Together they form the pre-distributed vector of asymmetric impact force on both legs.

[0036] Step 3: Generate independent shock wave propagation prediction parameter tables for the left and right legs; For the expected impact force amplitude of each leg in the pre-allocation vector of asymmetric impact force of the two legs, the expected time delay of the shock wave propagating from the foot to the ankle, knee and hip joints and the expected peak torque amplitude of each joint are pre-calculated using the structural equivalent longitudinal wave propagation velocity parameter of the leg and the equivalent impedance attenuation coefficient of each joint connection segment, and an independent shock wave propagation prediction parameter table for the left and right legs is generated.

[0037] It should be noted that the above-mentioned equivalent longitudinal wave propagation velocity parameters are pre-calibrated inherent parameters, characterizing the equivalent velocity of impact force propagating longitudinally in the leg structure, in meters per second. The above-mentioned equivalent impedance attenuation coefficient is the stiffness transfer ratio of each joint connection segment at the current joint angle, a dimensionless quantity, characterizing the degree of amplitude attenuation of the impact torque after passing through a connection structure. The expected time delay of each joint is equal to the equivalent structural path length from the foot to that joint (in meters) divided by the equivalent longitudinal wave propagation velocity parameter (in meters per second), with the result in seconds. The calculation method for the expected peak torque amplitude of each joint is as follows: Let the expected impact force amplitude at the foot of a certain leg be... (Unit: Newton), from foot to the 1st The path along the joint passes through a total of Segment connection structure, the first The equivalent impedance attenuation coefficient of the segment connection structure is (Dimensionless), then the expected peak torque amplitude of the joint is (Unit: Newton, representing the equivalent impact force amplitude), where Number the joints within a single leg. Numbers 1, 2, and 3 correspond to the ankle, knee, and hip joints, respectively. The numbers are for connecting structural segments. The value is obtained from a pre-calibrated stiffness transfer ratio table at the current joint angle.

[0038] It should be noted that the independent shock wave propagation prediction parameter table for each leg contains three sets of parameter records for each leg, corresponding to the ankle, knee, and hip joints respectively. Each set of parameter records includes the expected time delay value and the expected peak torque amplitude value for that joint. Due to the different expected impact force amplitudes for the left and right legs in the asymmetric impact force pre-assignment vector, the expected peak torque amplitudes for each joint in the independent shock wave propagation prediction parameter table for the left and right legs differ.

[0039] It should be noted that the pre-distribution vector of asymmetric impact force for both legs includes the expected impact force amplitude and the difference in expected contact time for each leg. In step 3, the expected impact force amplitude for each leg. These are respectively used as inputs for calculating the shock wave propagation prediction parameter table for the corresponding legs; the expected contact time difference. The advance amount used in step 4 to determine the offset of the ankle joint cushioning initiation time of the tilted leg is such that the cushioning initiation time of the tilted leg is advanced relative to the contralateral leg. Entering a compliant state.

[0040] Step 4: Generate a dual-leg differentiated cascaded buffer pre-scheduling scheme; Based on the independent shock wave propagation prediction parameter table for the left and right legs, with the optimization objective of evenly distributing the actual shock energy absorbed by each joint among all joints in the body, and with the torque limit of each joint, the upper limit of the joint angle travel, and the lower limit of the body roll angle recovery rate as constraints, a dynamic programming algorithm is used to calculate the buffer start-up offset and time-varying impedance parameter curve for each joint of each leg, generating a differentiated cascaded buffer pre-scheduling scheme for both legs.

[0041] It should be noted that the specific form of the above optimization problem is as follows: Let the six joints of the left and right legs be numbered as follows. ,in Number the joints, joints The actual energy absorbed during the buffering process is The objective function for uniform distribution is: ; in The total impact energy absorbed by the six joints. The peak amplitude of the expected torque corresponding to the independent shock wave propagation prediction parameter table for the left and right legs is denoted as . (Unit: Newton) From step 3, the corresponding leg and the corresponding joint position Value retrieval. By joint The time-varying impedance parameter curve and The buffer time is calculated by integral within the buffer time window, which is from the moment the joint starts buffering until the impact torque drops back to its initial value. The time corresponding to 20% of that.

[0042] The constraints include: ,in For joints The torque limit value; ,in For the joint during the cushioning process angular change For joints Upper limit of angle travel; ,in The angular velocity at which the torso rolls back to the neutral position after landing. This represents the lower bound of the roll angle recovery rate. The input to the dynamic programming algorithm is the expected time delay and the expected peak torque amplitude for the six joints of the left and right legs. Limit values ​​of torque at each joint Upper limit of joint angle travel and the lower limit of roll angle recovery rate The output of the dynamic programming algorithm is the buffer start-up offset and time-varying impedance parameter curves for each of the six joints.

[0043] It should be noted that the time-varying impedance parameter curves mentioned above include the stiffness and damping values ​​of each joint as a function of time during the buffering process. The buffering initiation time offset of the ankle joint on the tilted side is negative, indicating that it enters a compliant state earlier than the actual foot contact time; the buffering initiation time offset of the ankle joint on the opposite side is zero or positive. The initial damping value of the ankle joint on the tilted side is higher than that of the ankle joint on the opposite side to match the larger expected peak impact force.

[0044] In this embodiment, to ensure uniform distribution of impact energy among joints while also considering roll recovery, the dynamic programming algorithm divides its stages according to the chronological order of the shock wave arrival at each joint. In each stage, the maximum energy share that joint can absorb is determined based on the moment limit constraint of the corresponding joint, and any excess energy is transferred to the upstream joint of the next stage. The roll recovery rate constraint is satisfied by limiting the upper limit of the difference between the damping values ​​on the tilted side and the opposite side, ensuring that the angular velocity at which the torso recovers its roll angle to the neutral position after landing is not less than [a certain value]. .

[0045] Step 5: Preload the buffer scheme and perform real-time landing correction; Before landing, a differentiated cascaded buffering pre-scheduling scheme for both legs is preloaded into the parameter buffers of the impedance controllers of each joint. Upon impact, each joint sequentially enters a compliant buffering state according to the pre-configured activation time and time-varying impedance parameter curves. Simultaneously, real-time torque feedback values ​​from the torque sensors of each joint are acquired, and the deviation between the actual impact torque and the expected torque in the differentiated cascaded buffering pre-scheduling scheme for both legs is calculated. When the deviation exceeds a preset threshold, online correction is performed on the time-varying impedance parameter curves of upstream joints that have not yet activated buffering, redistributing the deviated energy to joints with sufficient capacity, and outputting the stable standing joint state after buffering is completed.

[0046] It should be noted that the aforementioned preloading occurs after the calculation in step 4 of the last control cycle during the air phase, but before the estimated landing time. The parameter buffers of each joint impedance controller have a double-buffered structure, supporting the simultaneous execution of current parameters and the receipt of updated parameter writes. This allows the dual-leg differentiated cascaded buffer pre-scheduling scheme updated in each control cycle during the air phase to continuously overwrite the writes until the landing time.

[0047] In this embodiment, to ensure effective buffering even when residual errors exist in the prediction during the air phase, an online correction process is added to step 5. The online correction process specifically includes: after obtaining the current torque feedback value for each activated buffer joint, calculating the difference between the current torque feedback value and the expected torque value at the corresponding moment in the dual-leg differentiated cascaded buffer pre-scheduling scheme as the torque deviation. When the absolute value of the torque deviation exceeds a preset threshold, it is determined that there is a residual error in the prediction, and the torque deviation value is transmitted to the time-varying impedance parameter curve corresponding to the upstream joint that has not yet activated buffering. If the torque deviation is positive, indicating that the actual impact is greater than expected, the initial damping value of the upstream joint is increased and its buffer activation time is advanced; if the torque deviation is negative, indicating that the actual impact is less than expected, the initial damping value of the upstream joint is decreased. The corrected parameters are written into the parameter buffer of the corresponding joint impedance controller, and the corrected parameters are used when the joint enters the buffering state at the corrected activation time.

[0048] This implementation establishes a real-time information coordination channel between the flight phase and the landing buffer phase through feedforward propagation of the landing attitude deviation prediction vector. During the flight phase, the roll and pitch deviations updated in each control cycle are converted into expected impact parameters for each joint via steps 2 and 3. Differential buffer parameters are pre-configured and loaded into each joint controller before landing. Because the tilted ankle joint has an earlier buffer activation time and a higher initial damping value before ground contact, the impedance controller is already in a compliant state when the impact reaches this joint, eliminating the scheduling lag caused by initiating propagation delay detection only after landing and preventing the ankle joint from bearing excessive load during the delay detection period. Independent shock wave propagation prediction parameter tables for the left and right legs match the buffering strategies of each joint in the left and right legs with the impact force amplitude they bear, avoiding the energy distribution imbalance caused by overload of the distal joints and insufficient buffering of the proximal joints under a unified buffering strategy. Real-time torque deviation correction after landing allows the differentiated cascaded buffering pre-scheduling scheme for both legs to maintain effective buffering through online compensation of upstream joints even when residual errors exist in the prediction, achieving coordinated control of flight sensing and landing buffering.

[0049] The following is an example of an application of the present invention, such as Figure 2-9 As shown, the implementation process is as follows: A bipedal parkour robot (model RB-07) performed a continuous obstacle course traversal task in an indoor obstacle training area. After completing a high platform traversal, RB-07 entered the flight phase. At this point, the inertial measurement unit detected a tendency for the torso to tilt to the right. The onboard real-time control system operates continuously with a control cycle of 5 milliseconds, requiring prediction and pre-scheduling during the flight phase and differentiated cascaded buffering upon landing.

[0050] After RB-07 takes off, the control system reads the trunk angular velocity output by the inertial measurement unit and the joint angular velocities output by the joint encoders in each control cycle. It then sums the angular velocity contributions of each body segment using the composite inertia matrix of the whole-body rigid-body tree structure to obtain the system's measured angular momentum three-dimensional vector. This vector is then subtracted component by component from the angular momentum reference value frozen at takeoff time to generate the angular momentum residual vector for the current cycle. As flight time progresses, the angular momentum residual vector sequence continuously accumulates. In the 18th control cycle of the flight phase (approximately 90 milliseconds from takeoff), the angular momentum residual sequence from takeoff to this cycle is taken, Z-score normalized, and input into a causal convolutional temporal prediction network. The network output, after denormalization, yields the landing attitude deviation prediction vector, as shown in Table 1.

[0051] Table 1. Landing attitude deviation prediction vector (output of the 18th control cycle)

[0052]

[0053] The roll angle deviation is positive, indicating that the RB-07 is expected to tilt its torso to the right by about 0.09 arcs upon landing, with its right leg tilted to the side.

[0054] The roll angle deviation output in step 1 Radius and pitch angle deviation The radius is used as input, combined with the parameters of the RB-07 kinematic model (hip joint width). Meters, lateral distance between the two feet Meters, total gravity (Newton), calculate the pre-assigned vector of asymmetric impact force on both legs.

[0055] The height difference between the left and right feet is Meters. The estimated current vertical falling velocity obtained by integrating with the inertial measurement unit. m / s, the time difference between contact between the left and right feet is 1 m / s The right foot (on the tilted side) is expected to touch the ground approximately 9 milliseconds earlier than the left foot.

[0056] Offset of the centroid projection in the left and right support regions Depend on and Calculated using the centroid position equation Meters (leaning to the right). The proportion of the normal impact force borne by the right leg (tilted side) is: The proportion of the left leg (opposite side) is The pre-distributed vectors of asymmetric impact force on both legs are shown in Table 2.

[0057] Table 2 Pre-distribution vector of asymmetric impact force on both legs

[0058]

[0059] The expected impact amplitude of the right leg is shown in Table 2. Newton, left leg Newton's time was used as the calculation input for the shock wave propagation prediction parameter table for each leg. The calibrated value of the equivalent longitudinal wave propagation velocity parameter of the RB-07 leg structure is... m / s. The equivalent structural path length from each joint to the foot and the equivalent impedance attenuation coefficient obtained at the current joint angle are shown in Table 3.

[0060] Table 3 RB-07 Leg Structure Parameters (at Current Joint Angle)

[0061]

[0062] The expected peak torque amplitude of each joint is as follows Calculation, where The expected normal impact force amplitude at the foot. For the first The equivalent impedance attenuation coefficient of each joint. Taking the right ankle joint as an example, Newton; knee joint Newton; hip joint Newtons. The results of calculations for each joint of the left leg using the same attenuation coefficient sequence with 306 Newtons as the base are summarized in Table 4.

[0063] Table 4 Prediction Parameters for Shock Wave Propagation in Left and Right Legs

[0064]

[0065] Table 4 clearly shows that the expected peak torque amplitude of each joint of the right leg is higher than that of the corresponding joint of the left leg, which is consistent with the pre-allocation result of the right leg bearing greater impact force in step 2.

[0066] The expected time delay and expected peak torque amplitude of the six joints are shown in Table 4. And the torque limits of each joint of RB-07 Upper limit of angle travel and lower limit of roll angle recovery rate Using radians per second as input, the dynamic programming algorithm unfolds stage by stage according to the order in which the shock wave arrives at each joint (ankle, knee, hip). In each stage, the maximum energy share that the current joint can absorb is determined, exceeding the target of uniform energy distribution. Partial transmission is made to the upstream joints. The roll angle recovery rate constraint is satisfied by limiting the difference between the initial damping values ​​of the right and left ankle joints to no more than a given upper limit. The algorithm outputs a bi-leg differentiated cascaded buffer pre-scheduling scheme, as shown in Table 5.

[0067] Table 5. Differentiated Cascaded Buffer Pre-scheduling Scheme for Two Legs (Key Parameters)

[0068]

[0069] The offset of the right ankle joint's buffer initiation time is -9.0 milliseconds, which is consistent with the value calculated in step 2. The "second-consistent" rating indicates that the right foot enters a compliant state 9 milliseconds before expected ground contact. The initial damping value of the right ankle joint, 185 N·s / radian, is higher than that of the left ankle joint, 151 N·s / radian, to match the larger expected peak impact force of the right leg.

[0070] After completing step 4 of the calculation in the last control cycle of the flight phase, the differentiated cascaded buffer pre-scheduling scheme for both legs is written into the parameter buffer of each joint impedance controller via a dual buffer structure, completing the pre-loading. The right ankle joint has entered a compliant buffer state according to the pre-configured parameters 9 milliseconds before the expected ground contact time, and the left ankle joint initiates buffering at the expected ground contact time.

[0071] After actual landing, the torque sensors at each joint continuously collect real-time torque feedback values. Taking the right ankle joint as an example, the expected peak torque amplitude is 307 Newtons, while the actual peak torque feedback is 341 Newtons, resulting in a torque deviation of [missing value]. The error exceeded the preset threshold (25 Newtons). The system determined that there was a residual error in the prediction and transferred this deviation to the time-varying impedance parameter curve of the right knee joint, increasing the initial value of the knee joint damping and advancing its buffer activation time. The corrected parameters were then written into the knee joint parameter buffer. The actual peak torque feedback value of the left ankle joint was 298 Newtons, deviating from the expected value of 251 Newtons. Newton also triggered an online correction, increasing the initial value of the left knee joint damping accordingly. The correction execution details are shown in Table 6.

[0072] Table 6 Comparison of key joint parameters before and after online correction.

[0073]

[0074] Note: The knee joint has not yet activated the buffer after the ankle joint correction is completed. "—" indicates that there is no measured value at this time, and the correction parameters have been written into the parameter buffer and will take effect when activated.

[0075] The entire data stream starts from the roll angle deviation output in step 1. Starting from the arc, step 2 converts it into an asymmetric impact force distribution of 374 Newtons for the right leg and 306 Newtons for the left leg, along with a 9-millisecond contact time difference. Step 3 converts this into the expected time delay and expected peak torque amplitude for each of the six joints. Step 4 forms a differentiated pre-scheduling scheme with the right ankle joint starting 9 milliseconds earlier and having a higher initial damping value than the left leg. Finally, step 5 completes pre-loading and compensates for knee joint parameters through online correction. There are clear numerical transmission relationships between the inputs and outputs of each step. The attitude deviation information acquired in the air phase completes the end-to-end configuration of differentiated buffer parameters before landing through the feedforward channel.

[0076] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for real-time cooperative control of a robot parkour motion, characterized by, include: During the flight phase, the angular momentum residual vector is calculated based on the torso angular velocity and joint angular velocity. The sequence of angular momentum residual vectors is input into the causal convolutional temporal prediction network to output the landing attitude deviation prediction vector. Based on the roll angle deviation and pitch angle deviation in the landing posture deviation prediction vector, the asymmetric impact force pre-distribution vector of the two legs is calculated in combination with the robot kinematics model. For the expected impact force amplitude of each leg in the pre-allocation vector of the asymmetric impact force of the two legs, based on the structural equivalent longitudinal wave propagation velocity parameter and the equivalent impedance attenuation coefficient, an independent shock wave propagation prediction parameter table for the left and right legs is generated. Based on the shock wave propagation prediction parameter table, a dynamic programming algorithm is used to calculate the buffer start-up time offset and time-varying impedance parameter curve for each joint, generating a differentiated cascaded buffer pre-scheduling scheme for the two legs. The pre-scheduling scheme is preloaded into the impedance controller of each joint. Upon landing, each joint enters a compliant buffer state according to the pre-configured parameters, and online correction is performed on the upstream joints that have not yet started buffering based on real-time torque feedback. 2.The real-time cooperative control method of robot parkour motion according to claim 1, characterized in that, The calculation method for the angular momentum residual vector includes: In each control cycle, based on the angular velocities of each joint and the trunk, the angular velocity contributions of each body segment are summed using the composite inertia matrix of the whole-body rigid tree structure to obtain the three-dimensional vector of the measured angular momentum of the system. The three-dimensional vector of the measured angular momentum of the system is then subtracted component by component from the reference value of angular momentum recorded at the take-off moment to obtain the residual vector of angular momentum for the current cycle.

3. The real-time collaborative control method for robot parkour movements according to claim 1, characterized in that, The causal convolutional temporal prediction network is a one-dimensional temporal convolutional network, whose convolutional kernels only apply operations to data at the current time and historical time. The input layer of the causal convolutional temporal prediction network receives a sequence of angular momentum residual vectors from the start time to the current control cycle. The input at each time step in the sequence is a three-dimensional angular momentum residual vector. The output layer is a fully connected layer that outputs a four-dimensional bias vector, which includes roll bias, pitch bias, roll velocity bias and pitch velocity bias. Before inputting the causal convolutional temporal prediction network, each component of the angular momentum residual vector sequence is normalized using Z-score, and the network output is denormalized to restore the physical dimensions.

4. The real-time collaborative control method for robot parkour movements according to claim 1, characterized in that, The pre-distribution vector of asymmetric impact force for both legs includes the expected contact time difference and the expected normal impact force amplitude for each leg; wherein, the calculation method for the expected contact time difference includes: Based on roll angle deviation and hip width Calculate the height difference between the left and right feet. Combined with the current vertical falling speed Calculate the time difference between the contact time of the left and right feet. .

5. The real-time collaborative control method for robot parkour movements according to claim 4, characterized in that, The method for distributing the expected normal impact force amplitude of each leg includes: The offset of the centroid projection in the left and right support regions is determined based on the pitch angle deviation and roll angle deviation. The proportion of the normal impact force borne by the inclined leg is calculated according to the moment balance relationship. The proportion of normal impact force borne by the opposite leg is ,in The lateral distance between the two feet. Due to roll angle deviation and pitch angle deviation The predicted normal impact force amplitude of each leg is calculated using the center-of-mass position equation in the robot's kinematic model; the predicted normal impact force amplitude of each leg is equal to the robot's total weight. Multiply by the corresponding proportion.

6. The real-time collaborative control method for robot parkour movements according to claim 1, characterized in that, The shock wave propagation prediction parameter table contains three sets of parameter records for each leg, corresponding to the ankle, knee, and hip joints respectively. Each set of parameter records includes the expected time delay value and the expected peak torque amplitude value; wherein: The expected time delay of each joint is equal to the equivalent structural path length from the foot to that joint divided by the equivalent longitudinal wave propagation velocity parameter of the structure; The expected peak torque amplitude of each joint is ,in The expected impact force amplitude at the foot of the leg. For the first The equivalent impedance attenuation coefficient of the segment connection structure at the current joint angle. Take 1, 2, and 3 in sequence, corresponding to the ankle joint, knee joint, and hip joint, respectively.

7. The real-time collaborative control method for robot parkour movements according to claim 1, characterized in that, The dynamic programming algorithm aims to uniformly distribute the actual impact energy absorbed by each joint across all joints in the body. The objective function is: ； in Number the joints. For joints The actual energy absorbed during the buffering process The total impact energy absorbed by the six joints; The constraints include: ,in For joints The expected peak torque amplitude, For joints The torque limit value; ,in For the joint during the cushioning process angular change This is the upper limit of the angle travel; ,in The angular velocity at which the torso rolls back to the neutral position after landing. This represents the lower limit of the recovery rate.

8. The real-time collaborative control method for robot parkour movements according to claim 1, characterized in that, The time-varying impedance parameter curve includes the stiffness and damping values ​​of each joint as a function of time during the buffering process. The offset of the cushioning initiation moment of the tilted leg ankle joint is negative, indicating that it enters a compliant state earlier than the actual foot contact moment; the initial damping value of the tilted leg ankle joint is higher than the initial damping value of the contralateral leg ankle joint.

9. The real-time collaborative control method for robot parkour movements according to claim 1, characterized in that, The online correction includes: At each joint where the buffer has been activated, the current torque feedback value is obtained, and the difference between the current torque feedback value and the expected torque value at the corresponding moment in the pre-scheduling scheme is calculated as the torque deviation. When the absolute value of the torque deviation exceeds a preset threshold, the torque deviation value is transmitted to the time-varying impedance parameter curve corresponding to the upstream joint that has not yet started the buffer; if the torque deviation is positive, the initial damping value of the upstream joint is increased and its buffer start time is advanced. If the torque deviation is negative, the initial damping value of the upstream joint is reduced; the corrected parameters are written into the parameter buffer of the corresponding joint impedance controller. The parameter buffer is a double-buffered structure that supports receiving updated parameters while the current parameters are being executed.

10. A real-time collaborative control system for robot parkour movements, used to execute the real-time collaborative control method for robot parkour movements according to any one of claims 1 to 9, characterized in that, include: The attitude deviation prediction module is used to calculate the angular momentum residual vector based on the torso angular velocity and joint angular velocity during the flight phase. The angular momentum residual vector sequence is input into the causal convolutional temporal prediction network to output the landing attitude deviation prediction vector. The impact force pre-distribution module is used to calculate the asymmetric impact force pre-distribution vector of the two legs based on the roll angle deviation and pitch angle deviation in the landing posture deviation prediction vector, combined with the robot kinematic model. The propagation prediction module is used to generate independent shock wave propagation prediction parameter tables for the left and right legs based on the structural equivalent longitudinal wave propagation velocity parameters and the equivalent impedance attenuation coefficient for the expected impact force amplitude of each leg. The buffer scheduling module is used to calculate the buffer start-up time offset and time-varying impedance parameter curve for each joint based on the shock wave propagation prediction parameter table and through dynamic programming algorithm, and generate a differentiated cascaded buffer pre-scheduling scheme for the two legs. The execution correction module is used to preload the pre-scheduling scheme into the impedance controller of each joint, control each joint to enter a compliant buffer state according to the pre-configured parameters when landing, and perform online correction on the upstream joints that have not yet started buffering based on real-time torque feedback.

Images