A target following control method and system for a legged humanoid robot

By employing a method of graded velocity shaping, phase-based pose alignment, and maintaining the target state using a world-fixed coordinate system, several problems in target following control of legged humanoid robots were solved, enabling the robot to achieve stable following and alignment with dynamic and static targets.

CN123114019APending Publication Date: 2026-11-17UQI TECH CO LTD
View PDF 9 Cites 0 Cited by

Patent Information

Application Number
CN202611326193.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-29
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies for target following control in legged humanoid robots suffer from several problems, including a lack of instruction adaptation between the following decision layer and the gait execution layer, a dual nature in speed command processing, failure to utilize omnidirectional translation capabilities, distortion in maintaining the target state using the body's relative coordinate system, the inclusion of the followed target body in the obstacle map, and observation jitter caused by tracking markers and gait fluctuations in target retrieval.

Method used

By employing graded velocity shaping and safe zero-speed maintenance, a phased pose alignment sub-process utilizing omnidirectional translation capability is implemented. The target state is maintained using a world fixed coordinate system, and target body cost elimination is set to achieve a five-level graded loss recovery sub-process, suppressing gait jitter and constructing a target following control system.

Benefits of technology

It enables the execution of speed commands for legged robots, eliminates stationary crawling, ensures that safety commands take effect immediately, improves the efficiency of close-range pose alignment, ensures accurate estimation of target state, avoids obstacle avoidance logic lock-up, and achieves stable dynamic target following and static target picking pose alignment.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123114019A_ABST
    Figure CN123114019A_ABST
Patent Text Reader

Abstract

This invention relates to the field of robot autonomous navigation and motion control technology, specifically a target following control method and system for a legged humanoid robot. It includes three coupled processing steps: graded velocity shaping and safe zero-velocity maintenance; a phase-based pose alignment sub-process based on omnidirectional translation capability; and a world-fixed coordinate system target state maintenance and five-level graded loss recovery sub-process. An executable velocity command for the legged platform is output through a three-level velocity shaping closed loop of error zeroing, amplitude enhancement, and acceleration / deceleration shaping. Static target pose alignment is completed by dividing multiple phases, and lateral errors are eliminated using omnidirectional lateral translation. The target state is maintained in the world-fixed coordinate system, and target body cost elimination and five-level graded loss recovery sub-processes are set. This invention achieves dynamic target distance-preserving following and static target object-picking pose alignment for a legged humanoid robot, balancing follow executability, alignment accuracy, and robustness in occluded scenarios. It has low computational complexity and is suitable for embedded real-time deployment.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of robot autonomous navigation and motion control technology, and in particular to a target following control method and system for a legged humanoid robot. Background Technology

[0002] Target following in mobile robots refers to the robot's real-time calculation and output of motion commands based on target observations provided by onboard sensing components, maintaining a desired relative positional relationship with the target being followed. A typical implementation involves the sensing components outputting the target's 3D bounding box, category, and tracking identifier. The controller selects and locks onto a target from this data, using the target's distance and azimuth as error parameters. A proportional-integral control law or an artificial potential field control law is then used to calculate the velocity command, which is then limited and sent to the chassis for execution.

[0003] The aforementioned framework is relatively mature on wheeled and tracked mobile platforms, but it will fail when directly adapted to legged humanoid robots due to differences in platform characteristics. Compared to wheeled platforms, legged humanoid robots face three unique constraints in following tasks: First, speed commands have an unexecutable lower dead zone. The lower-level gait controller of a legged platform needs to convert the body speed command into a foot landing point and swing leg trajectory. When the amplitude of the speed command is below a certain value, a valid stepping motion cannot be generated; the command is effectively swallowed by the lower layer, and the robot remains standing without displacement. Wheeled platforms, on the other hand, can have their drive wheels rotate continuously at any low speed, eliminating this constraint.

[0004] Secondly, it possesses omnidirectional translational capability and is not constrained by non-holonomic constraints. Legged humanoid robots can translate laterally in place without changing their orientation, while the lateral movement of differential wheel chassis can only be indirectly achieved through a combination of "turning-forward-turning" actions.

[0005] Third, the aircraft continuously undulates and sways during its gait cycle. The camera mounted on the aircraft shakes periodically accordingly, causing periodic perturbations in the observation of the center position, distance, and azimuth of the target bounding box, all coupled with the gait frequency.

[0006] The existing technology is as follows: (I) Target localization and following based on multi-source information fusion Patent CN122331545A discloses a robot target following method, apparatus, device, and storage medium. This solution acquires wireless positioning data, target image data, and target depth data of the target to be followed, determines the target position based on the fusion result of these three data points, and then controls the robot to perform following motion based on the target position, thereby improving the accuracy and stability of target position determination in multi-target, occluded, or complex environments.

[0007] The technical problem with this solution is that its improvements focus on the observation stage of "where is the target", but it does not address the execution stage of "how to convert position errors into motion commands that can be executed by the platform". The output follow motion commands may still fall within the speed range that the footed platform cannot execute. At the same time, its target positioning depends on the followed object carrying a wireless positioning device, which is not available in general following scenarios where the followed object does not carry such a device.

[0008] (II) Following based on human body part feature re-identification Patent CN121640027A discloses a robot target following method, apparatus, and device. This solution acquires an image of the target to be followed, detects the position information of the human body detection box, extracts feature information from preset positions such as the head, limbs, and center point of the human body based on human proportions to construct a feature vector, determines the target to be followed by comparing the similarity with the feature vector of the initial target, and then determines the robot's following parameters based on a proportional-integral-derivative controller, the robot's current state, and the predicted state of the target.

[0009] The technical problems with this solution are as follows: its contribution lies in reducing the computational complexity of target re-identification, but the control loop is still a conventional proportional-integral-derivative controller, whose output is a continuous velocity quantity. When the target approaches the desired distance, the error decreases and the output velocity also decreases, which happens to fall into the unexecutable lower dead zone of the legged platform, resulting in the robot not responding to small errors at all. In addition, this solution does not address the issue of the followed target itself being included in the obstacle map, nor does it provide a means for close-range object retrieval pose alignment.

[0010] (III) Composite robot following based on redundancy compensation of operating components Patent CN121552396A discloses a composite robot target following method and related equipment. This method obtains the actual distance and actual deviation angle from the target to the operating component, and determines the end effector pose control quantity of the operating component accordingly. Before exceeding the pose boundary, the positions of each joint motor are determined through inverse kinematics and the motion is executed. Simultaneously, the moving component is synchronously controlled based on the distance and deviation angle from the operating component to the moving component, thereby improving the speed and stability of the following response.

[0011] The technical problem with this solution is that the improvement in its following accuracy comes from the compensation of the following error of the moving component by the redundant degrees of freedom of the operating component (robotic arm). In essence, it transfers the insufficient control of the chassis to the robotic arm. Once the target exceeds the working space of the robotic arm, the compensation becomes ineffective, and the speed executability problem on the chassis side still exists. Moreover, this solution requires the robot to be equipped with an operating component that can be reversed in real time, which is not suitable for occasions where only chassis following is required.

[0012] (iv) Following in humanoid robot companion scenarios Patent CN120686839A discloses a method for a humanoid robot to follow and accompany a caregiver. This method uses cameras or depth sensors to identify and track the image and position of the caregiver, acquires environmental data using LiDAR and infrared sensors, employs the Dijkstra algorithm for path planning and obstacle avoidance, adjusts speed and direction according to the caregiver's movement to achieve real-time following while maintaining an appropriate distance, monitors the caregiver's vital signs, issues alarms in case of abnormalities, and provides voice interaction.

[0013] The technical problems with this solution are: its following control is summarized as "adjusting speed and direction according to the object's movement", without giving a shaping rule for the speed command, and therefore it cannot handle the dead zone of the speed of the legged platform; its obstacle avoidance uses the Dijkstra algorithm to perform a global search on the grid map, which requires high-frequency replanning in the following scenario where the target is continuously moving, resulting in large computational overhead and the path is prone to sudden changes between replanning; this solution also does not utilize the omnidirectional translation capability of the humanoid robot, nor does it address target retrieval after the tracking marker is switched.

[0014] (V) Obstacle Avoidance Based on Trajectory Prediction and Finite Time Domain Optimization Patent CN117420828A discloses a robot following and obstacle avoidance method and related equipment. This method acquires the motion state information and velocity change information of the pedestrian target being followed, predicts the future trajectory of the pedestrian target based on this information, and then determines the robot's desired position based on the future trajectory, thereby transforming the desired following into an iterative optimization problem in a finite time domain.

[0015] (vi) Compliant following based on virtual interactive force and admittance control Patent CN117193327A discloses a robot following obstacle avoidance method and related equipment. This method acquires the virtual interaction force between the robot and the human body to be followed, as well as the virtual potential field force between the robot and the obstacle. Based on these two factors, it calculates the robot's pose information after admittance control correction. Then, it calculates the control speed information from the corrected pose information to control the robot's movement. This integrates the artificial potential field method into the admittance control framework to improve the compliance and comfort of following.

[0016] (vii) Obstacle Avoidance Based on Velocity Sampling and Trajectory Selection Patent CN115509231A discloses a robot following obstacle avoidance method, device, and storage medium. This solution acquires surrounding environmental information during the following process, determines whether there are obstacles on the following trajectory, and if so, sequentially selects speed values ​​from the robot's speed range. Based on this, it calculates multiple predicted position coordinates within a preset time period to generate motion trajectories. Trajectories with a minimum distance to the obstacle greater than a preset threshold are retained as obstacle avoidance trajectories, and the one closest to the target position is selected as the optimal trajectory for execution.

[0017] The common technical problem with the solutions mentioned in items (v) to (vii) above is that all three focus their improvements on "how to generate a better desired trajectory or a more compliant desired velocity," while their output desired velocity is a real quantity defined over a continuous interval. Finite-time optimization outputs a tiny velocity that tends to zero near convergence, and the compliant characteristic of admittance control itself is designed with the output of a continuous small velocity as its goal. The candidate velocity set of the velocity sampling method also contains a large number of tiny velocities. On a legged platform, these tiny velocity commands are all unexecutable, and the robot will stagnate in a state of "issuing commands but not taking steps," and the error cannot converge. Subsequently, the error accumulates to the executable interval and then suddenly takes a step, which macroscopically manifests as a stop-and-go wobbling motion. All three also treat the followed target the same as ordinary obstacles: after the perception system writes the person being followed into the cost map, the attraction term pointing to the target and the repulsion term moving away from the target act on the same object. The closer the robot gets to the target, the stronger the deceleration becomes, until it is locked by its own obstacle avoidance logic outside the desired following distance. Furthermore, the target association of the three all depends on the continuity of the tracking identifier. However, after the occlusion ends, the tracker will generally assign a new identifier to the same object, at which point matching by identifier will inevitably fail.

[0018] (VIII) Control Scheme for Gait and Foot End Layer of Legged Robots Patent CN122469870A discloses a multimodal perception and terrain-adaptive gait control method and system for legged robots. This scheme acquires four-dimensional millimeter-wave radar point cloud data of the terrain in front of the robot, determines the pre-landing area, extracts terrain observation features, and obtains predicted values ​​for friction coefficient and stiffness. After the foot touches the ground, the actual friction coefficient and actual stiffness are obtained based on body perception data. A local map is used to perform spatiotemporal matching between the actual foot placement point and the previously pre-landing area to self-supervised update the prediction model. Based on the correction results, the terrain risk category is determined, gait parameters are adjusted, and the impedance parameters and friction constraints of the whole-body dynamic balance controller are updated.

[0019] Patent CN122469872A discloses a method for preventing foot slippage in quadruped robots operating on unstructured terrain. This method collects multi-dimensional force signals from the foot, leg joint motion signals, and inertial measurement signals from the robot body during the foot-to-ground contact phase. It then uses an equivalent dynamic model of the foot-ground contact to identify the equivalent ground stiffness and static friction coefficient online. A contact state vector is constructed to calculate local slip divergence trend indices to proactively predict slippage risks. When a risk occurs, the support polygon is dynamically contracted, and the friction cone constraint is reconstructed online. Finally, the joint torque and foot force commands are calculated.

[0020] The common technical problem with the two solutions mentioned above is that both aim to improve the gait generation and foot contact control of the foot-based platform, which is downstream of the body speed command. Their inputs are the already determined body speed command or gait parameters. Neither solution addresses how the body speed command itself is generated by the following error, thus failing to solve the problem of the speed command output by the upstream control law falling into the unexecutable range. On the contrary, it is precisely because of the physical constraint of the minimum executable step size in the gait layer that the upstream following control law must be adapted accordingly.

[0021] (ix) Summary of technical problems existing in the prior art Based on the above-mentioned existing technologies, the following technical shortcomings exist in the target following problem of legged humanoid robots: First, there is a lack of targeted instruction adaptation between the following decision-making layer and the legged gait execution layer. Existing following-type solutions are designed for wheeled platforms, assuming that speed commands are continuously executable across the entire range; existing legged-type solutions only optimize gait generation downstream of speed commands. The interface between the two—that is, how to shape the continuous control law output into speed commands that the legged platform can actually execute—is blank.

[0022] Second, the handling of speed commands is binary, with costs at both ends. If the control law does not have an error dead zone, the robot will continuously output tiny speeds near the target due to minor errors. These speed commands may be swallowed by the gait layer or cause repeated starts and stops, manifesting as wobbling and leg shaking. If a large dead zone is set on the control law side, the error within the dead zone is permanently ignored, and neither the following distance nor the alignment accuracy can converge to the desired value. There is a lack of a hierarchical shaping mechanism that "determines small errors as zero and raises effective errors to the executable lower limit."

[0023] Third, the omnidirectional translational capability of legged humanoid robots is not utilized. Existing following and pose alignment schemes use a "turn-forward" motion mode under non-holonomic constraints. When aligning the end effector at close range, repeated turning is required, which is not only time-consuming, but also changes the orientation of the robot with each turn, disrupting the established relative posture relationship between the robot and the target, which is detrimental to subsequent object-grabbing actions based on the robot's orientation.

[0024] Fourth, the target's state is maintained within the relative coordinate system of the machine. The machine's own translation and rotation are mixed into the estimation of the target's motion, resulting in distorted target velocity estimation and incorrect extrapolation direction of the lost position; when searching for the target by rotating in place, the target's historical trajectory recorded in the relative system rotates by the same angle as the machine, and the historical trajectory then becomes completely invalid.

[0025] Fifth, the target being followed is included in the obstacle map. The attraction pointing towards the target and the repulsion moving away from the target are applied to the same object, and the robot is locked in place by its own obstacle avoidance and safety logic near the desired following distance.

[0026] Recovering a lost target relies solely on tracking identifiers. After occlusion ends, trackers typically reassign identifiers to the same object, inevitably leading to identifier matching failures and ultimately resulting in a search timeout and termination of tracking. However, the followed object is often still within the robot's field of vision at this point.

[0027] Seventh, the observation jitter caused by gait fluctuations was not specifically suppressed. The jitter was amplified by the judgment threshold of the control mode or alignment stage, causing the system to repeatedly switch around the threshold, and the output speed command oscillated accordingly.

[0028] Therefore, a new technical solution is urgently needed to solve the above-mentioned technical problems. Summary of the Invention

[0029] The purpose of this invention is to overcome the problems of the prior art and provide a target following control method and system for legged humanoid robots. This method addresses the following issues in existing legged humanoid robot following tasks: lack of instruction adaptation interface between the following decision layer and gait execution layer; the binary nature of speed instruction processing failing to balance accuracy and executability; low alignment efficiency and posture disruption due to lack of omnidirectional translation; estimation distortion caused by maintaining the target state using the body's relative coordinate system; obstacle avoidance lock-up caused by including the followed target in the obstacle map; target retrieval relying solely on tracking markers; and control oscillations caused by gait jitter. This invention achieves dynamic target distance-maintaining following and static target object retrieval posture alignment for the legged humanoid robot, realizing a closed loop of "walking - accurate alignment - keeping up".

[0030] The above objectives are achieved through the following technical solutions: This invention provides a target following control method for a legged humanoid robot, comprising the following steps: Step S1: Obtain the target observation information output by the robot's perception, as well as the robot's body pose in the world fixed coordinate system; Step S2: Perform graded speed shaping and safety zero speed holding processing; first, based on the error dead zone, zero the error smaller than the tolerance is judged, then the executable lower limit is raised for non-zero speeds, then acceleration and deceleration amplitude shaping is performed, and finally, a second forced zeroing is applied after the safety trigger, and the three-degree-of-freedom body speed command executable by the leg motion control unit is output. Step S3: Select the control mode branch based on the target category; if the target is identified as a retrievable object, enter the phased pose alignment process, use the omnidirectional translation capability of the legged robot to eliminate errors by phase division, and perform zero-speed transition before switching between different motion modes; if the target is a dynamic follower, perform dynamic follower control. Step S4: Maintain the target state and historical trajectory in the world fixed coordinate system, and perform target body cost removal processing; when the target is detected to be lost, trigger the five-level graded loss recovery sub-process to retrieve the target; Step S5: Send the final speed command obtained after the above steps to the foot motion control unit.

[0031] As an optimization of this method, in step S2, during the shaping of graded speeds and the maintenance of safe zero speed, the lifting of the executable lower limit has a zero-keeping characteristic; only speeds greater than zero and less than the minimum executable amplitude of the corresponding channel are lifted to the minimum executable amplitude of that channel; when the speed is zero, it remains unchanged at zero; after the lifting, acceleration and deceleration amplitude shaping is performed, and the secondary forced zeroing for safe stopping is set after the acceleration and deceleration amplitude.

[0032] As an optimization of this method, the phased pose alignment sub-process in step S3 includes an orientation alignment phase, a forward approximation phase, a lateral alignment phase, a zero-velocity transition phase, and a stability determination phase; the phase transition adopts a hysteresis condition composed of an entry threshold and an exit threshold, and also needs to meet the continuous frame count condition; the lateral alignment phase outputs a lateral translation speed to eliminate lateral errors and keeps the body orientation unchanged throughout the process; when switching motion modes, it is necessary to enter the zero-velocity transition phase and wait for the speed of the previous mode to return to zero before starting the new mode.

[0033] As an optimization of this method, the five-level graded loss recovery sub-process in step S4 is executed sequentially: short-term hold, kinematic extrapolation, in-situ waiting, directional rotation search, and relocking of the same category; the matching conditions are relaxed only when the directional rotation search stage is in progress, allowing relocking of the target based on the target category and the position threshold in the world fixed coordinate system; in the non-search state, the target is strictly matched based on the tracking identifier.

[0034] As an optimization of this method, the target body cost culling in step S4 only takes effect in dynamic follow mode; with the target position in the world fixed coordinate system as the center and a preset culling radius, the potential field repulsion force calculation and forward safety gating traversal of the corresponding grid are skipped; the target body cost culling is turned off in pose alignment mode.

[0035] As an optimization of this method, a target observation smoothing sub-step is included between step S1 and step S2; the target observation smoothing sub-step includes: applying sliding window mean filtering to the forward and lateral coordinates of the target under the machine system; applying first-order low-pass filtering with the shortest angular distance to the target azimuth angle; applying first-order low-pass filtering to the target distance; and suppressing periodic observation jitter caused by gait fluctuations.

[0036] As an optimization of this method, the dynamic following control in step S3 is specifically as follows: when in dynamic following mode, guide point potential field obstacle avoidance control is adopted; guide points are selected according to the target's historical trajectory to generate attraction, and obstacle grids generate repulsion and tangential bypass forces; angle adaptive synthesis weights are configured, and the greater the target's azimuth angle deviates from the front, the higher the attraction weight; it also includes potential field local minima detection, and when the local minima is determined, time-limited alternating direction rotation in place is performed to escape from the obstacle.

[0037] As an optimization of this method, step S2 also includes steering speed reduction processing, specifically: when the forward linear velocity is greater than zero, the forward linear velocity is reduced according to the yaw rate amplitude ratio; the detection distance of the forward safety gate is taken as the maximum value among the fixed safety distance, the reaction braking distance, and the dynamic estimated braking distance.

[0038] Another aspect of the present invention provides a target following control system for a legged humanoid robot, used to implement the control method described in any of the above claims, comprising: The visual perception component is used to output the 3D bounding box, category label, confidence score, and tracking marker of the target within the field of view; Environmental perception components are used to construct local cost maps; The robot pose estimation component is used to output the robot's pose and velocity in a fixed world coordinate system. The foot motion control unit is used to receive three-degree-of-freedom body velocity commands and calculate the foot landing point and joint trajectory; An embedded computing unit is used to run the follow control program; the embedded computing unit internally deploys the following modules: The target candidate filtering module is used to filter out target candidates that do not meet the following conditions from the perceived output targets; The stable locking and jump rejection module is used to achieve stable target locking and reject abnormal jump observations; The target observation smoothing module is used to filter target observation data and suppress periodic jitter caused by gait fluctuations. The World Fixed Coordinate System Target State Estimation Module is used to calculate the target position and velocity in the world fixed coordinate system. The path memory and guide point module is used to record the target's historical trajectory and generate follow guide points; the target body cost elimination module is used to remove the grid corresponding to the target to be followed from the cost map repulsion force and safety judgment. The control mode arbitration module is used to select the corresponding control law based on the distance between the task and the target. The close-range proportional-integral control module is used to calculate the following speed command under close-range operating conditions; The phased pose alignment module is used to execute the phased pose alignment sub-process for the target object; The potential field obstacle avoidance control module is used to implement potential field obstacle avoidance following control based on the guide point and cost map; The speed limiting and steering speed reduction module is used to limit the control output speed and reduce the forward speed during sharp turns. The forward safety gating module is used to detect obstacles in front of the robot, trigger the forward speed to zero and set the safety stop sign; An executable lower limit lifting module is used to perform zero-preservation executable lower limit lifting processing on speed commands. The acceleration / deceleration limiting and safety zeroing module is used to perform acceleration / deceleration limiting and to force zeroing a second time after safety triggering. The graded loss recovery module is used to execute the five-level graded loss recovery sub-process to complete the target retrieval and disposal; The health gating and emergency stop module is used to monitor the health status of the system and output zero speed when an emergency stop or posture timeout occurs. The task session and status reporting module provides interfaces for starting and stopping tasks, and reports task running status and target information.

[0039] The modules in this system operate according to the following data flow and calling relationships: The visual perception component outputs to the target candidate screening module; the target candidate screening module outputs to the stability locking and jump rejection module; the stability locking and jump rejection module outputs to the target observation smoothing module and the hierarchical loss recovery module; the target observation smoothing module outputs to the world fixed frame target state estimation module; the body pose estimation component outputs to the world fixed frame target state estimation module; the world fixed frame target state estimation module outputs to the path memory and guide point module, the target body cost elimination module, and the hierarchical loss recovery module; the environmental perception component outputs to the target body cost elimination module and the potential field obstacle avoidance control module; the control mode arbitration module receives data from the path memory and guide point module. The outputs of the guide point module, the task session and status reporting module, and the hierarchical loss recovery module are sent to the near-range proportional-integral control module, the phased pose alignment module, and the potential field obstacle avoidance control module; the outputs of the near-range proportional-integral control module, the phased pose alignment module, and the potential field obstacle avoidance control module are sent to the speed limiting and steering deceleration module; the output of the speed limiting and steering deceleration module is sent to the forward safety gating module; the output of the forward safety gating module is sent to the executable lower limit raising module; the output of the executable lower limit raising module is sent to the acceleration / deceleration limiting and safety zeroing module; the output of the health gating and emergency stop module is sent to the acceleration / deceleration limiting and safety zeroing module; and the output of the acceleration / deceleration limiting and safety zeroing module is sent to the foot motion control unit.

[0040] This invention provides a target following control method and system for a legged humanoid robot. Through graded velocity shaping and safe zero-speed holding, it enables the execution of speed commands for the legged robot, eliminates in-place crawling, and ensures immediate effectiveness of safety commands. A phased pose alignment sub-process eliminates lateral errors without changing the robot's orientation, improving object-grabbing alignment efficiency. Maintaining the target state using a world-fixed coordinate system and setting target body cost culling ensures accurate target state estimation and avoids obstacle avoidance logic lock-up. A five-level graded loss recovery sub-process enables continuous following in occlusion and tracking marker switching scenarios. Multiple anti-jitter processing methods ensure stable control modes without oscillations. The entire method has low computational overhead and can be directly deployed on a robot's embedded controller. Compared to existing technologies, it offers the following specific advantages: First, it eliminates the control blind zone caused by the dead zone at the speed of the legged platform. For any error deemed valid by the control law, the corresponding speed command is raised to the platform's executable amplitude, and the robot no longer experiences the stagnant state of "issuing a command but not moving." Errors smaller than the tolerance are explicitly judged as zero, and the robot stands stably. The system only has two states: "clearly stationary" and "moving at an executable speed," with no invalid intermediate states between them, allowing the following distance to converge without creeping.

[0041] Second, the safe stopping command will not be altered by the lower limit raising stage. Because the raising rule strictly maintains a zero value, and the raising occurs after the preceding safety gating, the precise zero speed generated by the preceding safety gating and the error dead zone will not be raised into a motion command with an executable amplitude. This property is necessary for placing the lower limit raising after the preceding safety gating; if the order is reversed, the speed generated by the raising will bypass the safety decision.

[0042] Third, the safe stop takes effect immediately upon triggering. Since the safe zeroing occurs after the acceleration / deceleration limit, the deceleration limit will not cause the forward speed to decay to zero over several control cycles, thus avoiding the trailing effect of the limit mechanism on the safety response.

[0043] Fourth, the time required for close-range pose alignment is reduced while maintaining the aircraft's orientation. Lateral errors are eliminated directly by lateral translation, eliminating the need for a combination of "turning-forward-turning" movements. The aircraft's orientation remains unchanged throughout the entire lateral alignment process, providing a stable relative posture for subsequent object-grabbing actions based on the aircraft's orientation.

[0044] Fifth, the control mode and alignment phase do not tremble due to gait jitter. The target observation smoothing sub-step, double threshold hysteresis, and continuous entry frame count are superimposed to make the determination of mode switching and phase shift insensitive to gait periodic jitter, and the output speed no longer oscillates with jitter.

[0045] Sixth, the robot is not locked by its own obstacle avoidance logic at the desired following distance. Because the target being followed is eliminated in the potential field repulsion and forward safety judgment, the attraction and repulsion effects no longer point to the same object, and the robot can stably stop at the desired following distance.

[0046] Seventh, uninterrupted tracking even when occlusion and tracking marker switching occur. The target state is maintained in a fixed world coordinate system, ensuring the correct extrapolation direction during occlusion and preventing the historical trajectory from becoming invalid during rotational search. The relocking condition of adding a position threshold to the same category allows the tracker to still find the same object after the marker is reassigned, while the position threshold prevents locking onto other similar objects nearby.

[0047] Eighth, it has low computational overhead and is suitable for embedded deployment on the robot body. The core operations of this invention are several scalar comparisons, a sparse sampling traversal of the cost map, and a constant number of vector operations. It does not require offline training, nor does it require online solving of optimization problems or global graph search. It can run in real time on the robot body with limited computing power according to the control cycle. Attached Figure Description

[0048] Figure 1 This is a logic block diagram of a target following control system for a legged humanoid robot according to the present invention. Figure 2 This is a flowchart of the safety determination, candidate locking, and five-level graded loss recovery sub-process in the target following control method for a legged humanoid robot described in this invention; Figure 3 This is a flowchart of the state update, control law arbitration, and graded velocity shaping process in the target following control method for a legged humanoid robot described in this invention. Figure 4 This is a phase transition diagram of the phased pose alignment process in the target following control method for a legged humanoid robot described in this invention. Figure 5 This is a table of key parameter values ​​in the target following control method for a legged humanoid robot described in this invention. The values ​​are preferred values ​​in this embodiment and can be adjusted according to the overall size of the robot, the characteristics of the gait controller, and the characteristics of the working environment. Detailed Implementation

[0049] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. The described embodiments are merely some, not all, of the embodiments of the present invention. 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.

[0050] like Figures 1-3 As shown, this solution provides a target following control method and system for a legged humanoid robot. It improves the following control from "single control law directly outputting continuous speed" to a three-level speed shaping closed loop of "error zeroing - amplitude increase - acceleration and deceleration shaping". Based on the unique omnidirectional translation capability and gait jitter characteristics of the legged humanoid robot, it reconstructs the handling method of close-range pose alignment and target loss recovery.

[0051] This invention consists of three mutually coupled essential technical features: Firstly, there is graded velocity shaping and safe zero-speed maintenance. On the control law side, errors smaller than the tolerance are judged as zero using the error dead zone, ensuring the robot remains stationary. On the safety chain side, the non-zero velocity output by the control law is raised to the executable lower limit, ensuring that all issued commands can be executed by the gait layer. The raising stage strictly maintains the zero value, thus preventing the precise zero velocity generated by the safety gating and error dead zone from being raised into motion commands. After raising, acceleration and deceleration are applied for amplitude limiting and shaping, and after amplitude limiting, the channels with safety gating are forcibly zeroed a second time. The order of these four stages is an irreplaceable component of this feature.

[0052] Secondly, the phased pose alignment sub-process is based on omnidirectional translation capability. Close-range pose alignment is divided into four phases: orientation alignment phase, forward approach phase, lateral alignment phase, and stability determination phase. Lateral error is directly eliminated by lateral translation without changing the body orientation. Each phase is migrated using hysteresis composed of entry and exit thresholds, supplemented by continuous entry frame counting, to suppress phase jitter caused by gait tremors. Lateral translation and turning are set to be mutually exclusive. Before switching between the two, a zero-speed transition phase is passed to make the velocity of the previous mode zero, so as to avoid compound motion from disrupting gait balance.

[0053] Third, the world fixed coordinate system target status maintenance and five-level graded loss recovery sub-process. The target's position, velocity, and historical trajectory are maintained in the world fixed coordinate system and reprojected back to the machine coordinate system when in use. The target body being followed is eliminated in the cost map by radius and does not participate in the repulsion force and forward safety judgment. After the target is lost, it goes through five levels of handling in sequence: short-term holding, kinematic extrapolation, in-situ waiting, directional rotation search, and relocking of the same type.

[0054] The three elements form a closed loop of "being able to move, being able to align, and being able to keep up": the first element ensures that every speed command issued can be executed by the footed platform; the second element ensures that close-range errors can be eliminated efficiently without affecting the orientation of the aircraft; and the third element ensures that the target is not dropped when the target is obscured or the tracking marker is switched.

[0055] The hardware requirements for implementing this invention are: (1) A set of visual perception components is installed on the head of the robot body to output the three-dimensional bounding box, category label, confidence score and tracking mark of each target in the field of vision; (2) A set of environmental perception components, which can be lidar or depth camera, used to construct a local cost map around the robot; (3) One body pose estimation component, used to output the robot's pose and velocity in the world fixed coordinate system; (4) A set of foot motion control unit, which receives the three-degree-of-freedom body velocity command consisting of forward linear velocity, lateral linear velocity and yaw angular velocity, and calculates it into the landing point and joint trajectory of the two legs; the unit has a minimum executable speed amplitude for each of the above three channels, and commands below the amplitude cannot generate effective steps; (5) An embedded computing unit for running the follow control program of the present invention.

[0056] The output of the visual sensing component needs to be pre-transformed to the machine base coordinate system through extrinsic parameter calibration, or it can be directly output by the visual sensing component in the machine base coordinate system. A coordinate transformation relationship that can be queried in real time must exist between the local cost map and the machine base coordinate system. The extrinsic parameter calibration transformation of the sensors is a well-known technique in the field of robot perception, and this specification will not elaborate on the specific calibration process. The system uses configuration items to distinguish whether the sensing output coordinate system and the machine base coordinate system are consistent, and automatically inserts an extrinsic parameter transformation step when they are inconsistent, allowing the same method to adapt to different overall sensor arrangements.

[0057] As one embodiment of this solution, in this embodiment, a legged humanoid robot performs companionship and object retrieval tasks in an indoor environment. The robot's head is equipped with a set of visual perception components, which outputs the 3D bounding boxes, category labels, confidence scores, and tracking markers for each target within its field of view. The robot body is equipped with a set of environmental perception components, used to construct a local cost map centered on the robot body. The robot body pose estimation component outputs the robot's pose and velocity in a fixed world coordinate system. The legged motion control unit receives three-degree-of-freedom velocity commands consisting of forward linear velocity, lateral linear velocity, and yaw angular velocity. The embedded computing unit runs the follow-up control program at a cycle of 50 Hz.

[0058] The threshold parameters described in this embodiment are calibrated based on the gait characteristics of the legged humanoid robot and the sensor noise level. Those skilled in the art can select and calibrate within the adjustable parameter range described in this application based on the actual working conditions of the whole machine hardware.

[0059] The tasks are divided into two categories: Task A is dynamic following, where the target being followed is category A (human pedestrian targets), and the robot needs to maintain the desired following distance and avoid obstacles along the way; Task B is static object retrieval pose alignment, where the target is category B (retrievable container targets), and the robot needs to servo the target to the relative pose window that the robotic arm can grasp and report that it can be grasped. The system automatically distinguishes between the two types of tasks based on whether the target category assigned by the task belongs to the preset set of retrievable object categories.

[0060] The working environment includes three typical working conditions: Condition X is when the object being followed walks in a straight line and turns a right angle; Condition Y is when the object being followed passes behind a column or door frame and is briefly obstructed; Condition Z is when the target object is placed on the edge of a table and the robot needs to get close and align with it.

[0061] like Figure 1 The diagram shown is a logical block diagram of the system composition of the present invention. The system includes the visual perception component, the body pose estimation component, and the environment perception component. The visual perception component outputs a signal to the target candidate filtering module. The target candidate filtering module filters out targets from all targets output by the perception component that are unlikely to become the follower, based on hard criteria. This module filters based on whether the category label matches the task specification, whether the confidence score is higher than the lower limit, whether the forward distance is greater than the minimum value, and whether the planar distance is less than the maximum value corresponding to the current control mode, selecting the closest candidate from the remaining candidates. Its output is then sent to the stable locking and jump rejection module.

[0062] The stable locking and jump rejection module prevents false detections of a single frame as a target and rejects abnormal observations after locking. This module requires that the tracking identifier, category, position, and angle remain continuous across multiple consecutive frames for the same candidate target, and that the confidence level does not significantly decrease before formal locking. After locking, if the position or angle jump of the same target relative to the original observation of the previous frame exceeds a threshold, the current frame is treated as if the target was not observed. Here, the original observation, rather than the smoothed result, is used as the jump comparison benchmark. This avoids the inherent lag in the smoothing process being misjudged as a target jump, thus preventing false rejections when the followed object is moving normally and rapidly. Its output is fed into the target observation smoothing module and the hierarchical loss recovery module.

[0063] The target observation smoothing module executes a target observation smoothing sub-step to suppress periodic jitter introduced by the gait fluctuations of the foot-based platform in target observation. This module applies a sliding window mean to the forward and lateral coordinates of the target in the body coordinate system, applies a first-order low-pass filter to the target distance, and applies a first-order low-pass filter to the target azimuth using the shortest angular distance. Using the shortest angular distance instead of direct interpolation avoids the low-pass filter from circling in the opposite direction when the azimuth is around ±180 degrees. Its output is fed into the world fixed-frame target state estimation module and various control modules.

[0064] The world-fixed coordinate system target state estimation module provides the target position and velocity, unaffected by the robot's own motion. This module transforms the target observation from the body coordinate system to the world-fixed coordinate system, differs from the previous position to obtain the instantaneous velocity, and then applies a first-order low-pass filter and amplitude limiting to obtain the target velocity estimate. This module forms the common foundation for path memory, kinematic extrapolation, and class-based relocking functions. Its output is fed into the path memory and guide point module, the target body cost elimination module, and the hierarchical loss recovery module.

[0065] The path memory and guide point module enables the robot to travel along the path already taken by the object being followed, rather than approaching it in a straight line. This module records a historical point in the world fixed coordinate system each time the target moves beyond the minimum sampling interval, and discards old points according to their duration. During control, it selects the first historical point closer to the target than the desired following distance as the guide point, following the historical trajectory from earliest to nearest. Each historical point is reprojected back to the current robot coordinate system during use. Its output is sent to the control mode arbitration module.

[0066] The target cost elimination module performs target cost elimination processing to prevent the followed target from being treated as an obstacle. This module uses the estimated target position in a fixed world coordinate system as the center and a preset elimination radius as the range. During grid traversal involving potential field repulsion force accumulation and forward safety gating, it skips grids falling within this range. This module is only effective in dynamic following mode and not in pose alignment mode because structures such as table edges adjacent to the target object in the object retrieval scenario must still trigger a safe stop. Its output is sent to the potential field obstacle avoidance control module and the forward safety gating module, respectively.

[0067] The control mode arbitration module selects the control law based on the task category and the distance to the current target, preventing repeated oscillations around the switching threshold. When the target category specified by the task belongs to the accessible object category, this module selects the phased pose alignment module to execute the phased pose alignment sub-process; otherwise, it selects between the near-range proportional-integral control module and the potential field obstacle avoidance control module based on the target distance, with different distance thresholds used for entry and exit to create hysteresis. Its output is the selected active control law.

[0068] The close-range proportional-integral control module is designed to maintain the following distance and orientation using a simple and stable control law when the target is close and the obstacle risk is low. This module uses the difference between the target distance and the desired following distance as the linear velocity error and the target azimuth angle as the angular velocity error. Both are first checked for zero error in the error dead zone, and then the speed command is calculated using the proportional and integral terms.

[0069] The potential field obstacle avoidance control module is used to perform guide point potential field obstacle avoidance control in dynamic following control. Its function is to balance approaching the target and avoiding obstacles when the target is far away or when it is necessary to bypass obstacles. This module generates an attractive force using the guide point and generates radial repulsive force and tangential bypass force using grids in the cost map that exceed the occupancy threshold and are located within the influence radius. It adaptively adjusts the combined weight of the attraction and repulsion terms based on the target azimuth angle, so that the robot prioritizes turning towards the target when the target is significantly deviated from direct front. This module also has a potential field local minimum detection: when the output velocity is low for a long time and the target is not within close range, it is determined to be trapped in a local minimum and switches to a timed alternating direction in-situ rotation to escape the obstacle. The rotation direction is reversed with each trigger, that is, if the current trigger is clockwise rotation, the next trigger will be counterclockwise rotation, to avoid the robot repeatedly hitting the same obstacle. During the escape, the forward linear velocity is set to zero.

[0070] The phased pose alignment module is used to execute the phased pose alignment sub-process. Its function is to use the omnidirectional translation capability of the legged platform to servo the static target into the relative pose window that the robotic arm can grasp.

[0071] The speed limiting and steering deceleration module constrains the control law output to the platform's allowable range and reduces forward speed during sharp turns. This module limits the speed of each of the three channels and reduces the forward speed proportionally to the angular velocity amplitude when the forward speed is positive. Its output is sent to the forward safety gating module.

[0072] The forward safety gating module prevents the robot from moving forward when an obstacle is in front of it. The detection distance of this module is the greater of a fixed safety distance and a braking distance estimated based on the current speed, and the detection width is a preset value. When there is an occupied grid within the detection area, or when the cost map and coordinate transformation are unavailable, the forward speed is set to zero and a safety stop sign is activated. Its output is sent to the executable lower limit lifting module, and the safety stop sign is also sent to the acceleration / deceleration limiting and safety zeroing module.

[0073] The executable lower limit lifting module eliminates the control blind band caused by the dead zone at the speed of the legged platform. This module determines the speed of each of the three channels: if the absolute value of the speed in that channel is greater than zero and less than the minimum executable amplitude of that channel, it raises the speed to that amplitude while maintaining the original sign; if the speed of that channel is exactly zero, it remains zero without any modification. This module is a key component that distinguishes this invention from the following scheme for wheeled platforms, and its output is sent to the acceleration / deceleration limiting and safety zeroing module.

[0074] The acceleration / deceleration limiting and safety zeroing module ensures that the speed command after acceleration does not impact the gait controller in a step-like manner, and that the safety stop command is not trailed by the limiting circuit. This module first limits the change in speed in the current cycle relative to the previous cycle based on the upper limits of acceleration and deceleration for each channel; then it checks the safety stop sign and emergency stop status of the forward safety gate; if any is set, the corresponding channel is forcibly zeroed a second time. Placing safety zeroing after acceleration / deceleration limiting prevents the upper limit of deceleration from causing the forward speed to decay to zero over several control cycles, thus achieving zero-speed output immediately upon triggering the safety gate. Its output is the final speed command sent to the foot motion control unit.

[0075] The hierarchical loss recovery module executes a five-level hierarchical loss recovery sub-process, which maintains tracking when the target is temporarily invisible and abandons it in an orderly manner when it is temporarily invisible. This module handles the loss in five stages sequentially: short-term hold, kinematic extrapolation, in-situ waiting, directional rotation search, and relocking of the same type. Its output is sent to the control mode arbitration module and the task session and status reporting module.

[0076] The health gating and emergency stop module is designed to stop the system unconditionally when the system status is unreliable or when an emergency stop command is received. This module monitors the freshness of the body pose estimation update and the emergency stop signal; any abnormality triggers a zero-speed output and terminates the current following task.

[0077] The task session and status reporting module provides interfaces for starting, preempting, and canceling follow-up tasks, and reports the current status, target information, and alignment error in real time.

[0078] like Figure 2 The diagram shows a flowchart of the safety determination, candidate locking, and five-level graded loss recovery sub-process in the following control method of this invention. After the process begins, the perceived target and the machine's pose are received. First, an emergency stop or positioning timeout is determined. If the determination is successful, zero-speed output is performed and the task ends; otherwise, candidate screening and stable locking are executed. It is then determined whether a valid target exists in the current frame. If no valid target exists, the five-level graded loss recovery sub-process is entered. It is then determined whether the target is within the hold or extrapolation window. If so, the target position is extrapolated according to the world fixed coordinate system velocity; if not, the target waits in place and then rotates to search. Next, it is determined whether a target of the same category falls within the prediction threshold. If the search timeout occurs, the task ends; if the conditions are met, the target is relocked and the smoother is reset. The process then proceeds to... Figure 3 .

[0079] like Figure 3 The diagram shown is a flowchart of the state update, control law arbitration, and graded velocity shaping processes in the following control method of the present invention. Figure 2Output: Execute the target observation smoothing sub-step and update the world fixed coordinate system state, record historical trajectories, and perform target body cost culling; determine if the target belongs to the accessible object category; if so, execute the phased pose alignment sub-process; if not, determine if the distance is less than the proximity threshold; if the condition is met, execute the proximity proportional integral calculation; otherwise, execute the guide point potential field obstacle avoidance control; then uniformly enter the speed limiting and steering deceleration, execute the forward safety gating and set the safety sign; raise the non-zero channel to the executable lower limit and execute acceleration / deceleration limiting; determine if the safety stop sign is set; if set, the corresponding channel is forcibly zeroed twice; if not set, retain the limiting result, issue speed commands and report the status.

[0080] like Figure 4 The diagram shows the phase transition of the phased pose alignment sub-process of this invention. The phases include the orientation alignment phase, forward approximation phase, lateral alignment phase, zero-velocity transition phase, and stability determination phase. If the angle error exceeds the entry threshold for multiple consecutive frames, the system transitions to the orientation alignment phase; if the angle error drops below the exit threshold, the system exits the orientation alignment phase. If the forward error exceeds the tolerance, the system enters the forward approximation phase; if the forward error falls within the tolerance, or if the forward distance is less than the priority threshold while the lateral error exceeds the threshold, the system enters the lateral alignment phase. Switching between the lateral alignment phase and the forward approximation phase must pass through the zero-velocity transition phase. When both the lateral and forward errors fall within the tolerance, the system enters the stability determination phase; if the number of consecutive stable frames reaches the target, the alignment is determined to be complete and can be reported for capture; if any error exceeds the tolerance again, the system returns to the corresponding phase. The zero-velocity transition phase is a forced transition between lateral translation, turning, and forward movement. Before switching motion modes, the velocity of the previous mode is reset to zero to avoid compound movements disrupting gait balance.

[0081] like Figure 2-5 As shown, the specific steps of the laser mapping method for a wheeled humanoid robot chassis provided in this embodiment are as follows: Step S1: Obtain the target observation information output by the robot's perception, as well as the robot's body pose in the world fixed coordinate system.

[0082] In this step, a safety and health check is performed first: it checks whether the emergency stop signal is set and whether the most recent update of the aircraft's pose estimation has exceeded the pose timeout threshold. If either condition is met, a zero-speed command is immediately output, the current following task is terminated, and a safe stop is reported externally. No further steps are executed in this cycle.

[0083] Then, target candidate screening and stable locking are performed: For each target output by the sensing unit in this frame, the following checks are performed sequentially: whether its tracking identifier is valid; whether its category label matches the category specified in the task; whether its confidence score is not lower than the lower confidence limit; whether its forward coordinate is not less than the minimum forward distance; and whether its planar distance is not greater than the maximum sensing distance corresponding to the current task. If a locked target already exists, only those with the same locking identifier are accepted, and their position jump and angle jump relative to the original observation of the previous frame are further checked to ensure that they do not exceed the jump threshold. Among the candidates that pass all checks, the one with the smallest planar distance is selected as the best candidate for this frame.

[0084] If the target has not yet been locked, the best candidate must maintain the same tracking identifier and category for 3 consecutive frames, with a position jump of no more than 0.55 meters, an angle jump of no more than 0.55 radians, and a confidence drop of no more than 0.80 before it can be officially locked; if the link is broken in any frame, the candidate in that frame will be the new target to be locked and the count will be restarted.

[0085] The target observation smoothing sub-step is performed after step S1 and before step S2: For the target in the body coordinate system, the average of the sliding window values ​​of the most recent 5 frames is taken for both the forward and lateral coordinates. The target distance is applied using Equation (1). The target azimuth is applied using Equation (2) in the form of the shortest angular distance. When the target changes, all smoother states are cleared to avoid the historical trail of the old target from polluting the first frame output of the new target.

[0086] , In the formula, The mean of the forward coordinates of the target during the k-th control cycle is expressed in meters; the horizontal coordinate is... The window mean is calculated in the same way. The reason for using the window mean instead of the low-pass filter is that the gait fluctuations of the foot platform introduce a periodic disturbance with an approximate zero mean in the target coordinate observation. When the window length covers one or two gait cycles, the mean calculation can directly cancel out this disturbance, and does not introduce the phase lag inherent in the low-pass filter that is related to the disturbance amplitude.

[0087] , In equation (2), This indicates that the angle is normalized to... Operations on intervals; This represents the low-pass coefficient; a higher value indicates greater confidence in the current observation and faster response to steering maneuvers, but also weaker jitter resistance. Target distance. according to Calculation. Equation (2) first takes the shortest angular distance for the angle difference and then performs a low-pass calculation. Its function is to make the azimuth angle positive and negative when the angle is in the positive and negative directions. When crossing nearby, the low-pass circuit approaches along the shortest direction instead of detouring in the opposite direction.

[0088] Step S2: Perform graded speed shaping and safe zero-speed holding processing.

[0089] Before performing this step, first complete the target state update and path memorization in the world fixed coordinate system: The smoothed target position is transformed from the body coordinate system to the world fixed coordinate system, and the difference between it and the previous effective observation position is calculated according to equation (3). After passing through a first-order low-pass filter and amplitude limiting, the target velocity estimate in the world fixed coordinate system is obtained. If the interval between the current and previous effective observations is less than 1 millisecond or greater than 1 second, the velocity estimate is not updated in this case to avoid divergence of the difference results when the interval is abnormal.

[0090] , In equation (3), This means restricting the magnitude of vector a to no more than [a certain value]. Saturation operations with unchanged direction; Indicates the speed low-pass coefficient; This represents the upper limit of the target velocity estimation amplitude. Its function is to suppress the situation where a perceived abnormal jump is amplified to a maximum velocity after differential processing, thus causing the extrapolated position to fly out. The key to equation (3) is... Defined in a fixed world coordinate system rather than a body coordinate system. If differences are made within the body coordinate system, the robot's own translations and rotations will be included in the target velocity: the robot's velocity... When moving forward, the coordinate difference result of the stationary target in the body coordinate system is: According to this extrapolation, the target position will move behind the robot, which is contrary to the fact; when the robot rotates in place to search, the target's historical trajectory recorded in the robot's coordinate system will rotate as a whole and become invalid.

[0091] Then, it is determined whether the target's movement distance relative to the most recent historical point has reached the minimum sampling interval of 0.1 meters. If it has, a historical point is added in the world fixed coordinate system; historical points with a recording duration of more than 5 seconds are eliminated.

[0092] Step S3: Select the control mode branch based on the target category; if the target is identified as a retrievable object, enter the phased pose alignment process, use the omnidirectional translation capability of the legged robot to eliminate errors by phase division, and perform zero-speed transition before switching between different motion modes; if the target is a dynamic follower, perform dynamic follower control.

[0093] Dynamic follow control is divided into two modes: close-range proportional-integral solution and guide point potential field obstacle avoidance control. ① Solving for proportional integrals at close range: The forward linear velocity and yaw rate are calculated according to formula (6): First, the distance error is judged to be zero by the larger of the distance tolerance and the near target distance dead zone, and the azimuth error is judged to be zero by the small angle dead zone; if the distance error and the azimuth error fall into the near target dead zone at the same time, then both are judged to be zero; then the speed command is calculated by the proportional term and the integral term, and the lateral linear velocity is set to zero.

[0094] , , In equation (6), This means restricting 'a' to the interval. Truncation operation; Indicates error dead zone operation, when Take zero if it is zero, otherwise take zero. ; This represents the upper limit of the yaw angular velocity, in radians per second. The lower bound of the forward linear velocity cutoff is zero, indicating that the proximity control law only allows forward movement and not backward movement. The error dead zone in the formula... The execution lower limit elevation in equation (9) constitutes the two ends of the graded speed shaping in this invention. The two act on different quantities and are indispensable: the former acts on the error, and its function is to determine "whether it is worth moving", judging that errors smaller than the tolerance do not require a response; the latter acts on the speed, and its function is to ensure that "once it is decided to move, it can move". If only the former is present, small errors outside the dead zone will still generate speed commands that fall within the non-executable range; if only the latter is present, any tiny error will be elevated into a motion command with an executable amplitude, and the robot will continuously start and stop near the target.

[0095] ② Guiding point potential field obstacle avoidance control: Search for the first historical point in the world fixed coordinate system from the earliest to the nearest after reprojection back to the body coordinate system. The distance between the robot and the historical point is not less than the expected following distance. Calculate the attraction force according to formula (4). If no suitable historical point is found, the current position of the target is used instead.

[0096] , In equation (4), This indicates the coordinates of the selected guide point in the body coordinate system, in meters; Indicates the sequence number of the historical trajectory point; Indicates the first The recorded moments of each historical point are in seconds. This represents the coordinates of the j-th historical point after it has been reprojected from the world fixed coordinate system back to the current machine coordinate system, in meters; The minimum distance limit of the guide point is expressed in meters. Equation (4) means that among all historical trajectory points, the one with the earliest recorded time and a current distance from the robot that is not less than the expected following distance is selected as the guide point. The earliest point is selected instead of the closest point so that the robot can travel along the route taken by the object being followed. The point-by-point reprojection before each use is used to ensure that the route is not affected by the robot's own movement.

[0097] Traverse the local cost map for grids whose occupancy value is not lower than the occupancy threshold, are not marked, and are within 0.01 meters to the radius of influence of the robot. Accumulate the radial repulsion force and tangential circumduction force according to Equation (5) and limit the amplitude of the repulsion resultant force. Then synthesize the total potential force according to the angle adaptive weight in Equation (5). Calculate the yaw rate with the direction of the total potential force through proportional gain. Calculate the forward linear velocity in segments based on the target distance: take the maximum linear velocity when the distance is not less than the full speed distance; reduce the speed according to the 2.5 power of the ratio of the two when the distance is between the stopping distance and the full speed distance; and set the speed to zero when the distance is not greater than the stopping distance.

[0098] , In equation (5), , This represents the remaining amount after deducting the expected following distance from the coordinates of the guide point; By each effective obstacle grid The weighted sum is accumulated along the direction away from the grid and then limited by the modulus. By each effective obstacle grid The normal direction along the repulsion direction of the grid is accumulated, and its sign is taken as the one that makes the angle between the tangential bypass direction and the attraction direction acute; G represents the angle adaptive synthesis weight. In equation (5), G increases with the increase of the absolute value of the target azimuth angle. Its function is: when the target is seriously deviated from the front of the robot, the weight of the attraction term is increased, so that the robot turns to the target first and is not continuously pushed away by the repulsion force of the side obstacle; when the target is basically in front, the weight of the attraction term is reduced, so that the obstacle avoidance force dominates and the robot can bypass it.

[0099] If the calculated forward linear velocity and yaw angular velocity are both below the local minimum threshold for 1.5 seconds for 1 consecutive seconds, and the target distance is not within the close range, then it is determined that the target is trapped in a local minimum potential field and enters a timed alternating direction rotation to escape the trap for 1.0 seconds. The rotation direction is reversed each time it is triggered, and the forward linear velocity is set to zero during the escape.

[0100] Phased pose alignment process solution: The forward and lateral errors are obtained by subtracting the desired forward and lateral positions of the target from its forward and lateral coordinates in the body coordinate system, respectively; the target's azimuth angle is used as the angular error. Figure 4The migration conditions determine the phase of the current cycle: when the absolute value of the angle error exceeds the rotation entry threshold for 3 consecutive frames, the phase migrates into the orientation alignment phase; when the absolute value of the angle error drops below the rotation exit threshold, the phase migrates out. When the forward error exceeds the forward tolerance, the phase is in the forward approximation phase. When the forward error has entered the tolerance and the lateral error exceeds the lateral tolerance, or when the forward distance is less than the lateral priority threshold and the lateral error exceeds the lateral priority entry threshold, the phase migrates into the lateral alignment phase. During the lateral alignment period, when the forward error exceeds the sum of the forward tolerance and the backoff redundancy, the phase migrates back to the forward approximation phase.

[0101] The velocity calculation for each phase is as follows: the orientation alignment phase only outputs the yaw rate, which is the value of the angle error multiplied by the angle gain and then limited; the forward approach phase outputs the forward linear velocity proportional to the forward error, and when the absolute value of the lateral error exceeds the auxiliary correction threshold, a small angular velocity not exceeding the upper limit of the auxiliary correction angular velocity is superimposed; the lateral alignment phase outputs the lateral linear velocity according to formula (11), and the forward linear velocity and yaw rate are both set to zero; the stability determination phase does not output any velocity.

[0102] , In equation (11), This represents the lateral alignment error, expressed in meters. Its value is the target lateral coordinate minus the desired grab lateral position. The lateral linear velocity proportional gain is represented by Equation (11). The lateral linear velocity output by Equation (9) must also be raised by the lower limit of the executable value, because there is also a minimum executable amplitude for lateral translation on the foot platform, and this amplitude is usually higher than the corresponding amplitude of forward motion.

[0103] When phase shifting involves switching between lateral translation and non-lateral motion, a zero-velocity transition phase is first entered: the velocity of the new mode is not output in this cycle, and the new mode is started in the next cycle only after the velocity command of the previous mode decays to zero.

[0104] When all three conditions are met simultaneously—the absolute value of the forward error not exceeding the forward tolerance, the absolute value of the lateral error not exceeding the lateral tolerance, and the forward coordinate not less than the minimum effective forward distance—the stable frame count is accumulated; if any one condition is not met, the count is reset to zero. When the stable frame count reaches 10 frames, alignment is determined to be complete, zero speed is output, and the data is reported to be ready for capture.

[0105] Step S4: Maintain the target status and historical trajectory in the world fixed coordinate system and perform target body cost removal processing; when the target is detected to be lost, trigger the five-level graded loss recovery sub-process to retrieve the target.

[0106] The target cost elimination process involves marking grids in the local cost map that fall within a 0.6-meter radius, centered on the estimated target position in the world fixed coordinate system. These marked grids are skipped in subsequent potential field repulsion accumulation and forward safety gating traversal. This process is only performed in Task A (dynamic following); it is not performed in Task B (pose alignment).

[0107] The five-level graded loss recovery sub-process: When step S1 fails to obtain a valid target, it is processed according to the following five levels in sequence, and the processing result is sent back to step S3: Level 1, short-term hold. If the time since the most recent valid observation does not exceed 0.8 seconds of the lost hold time, the control continues to use the previous valid target. Its function is to filter out one or two frames of transient detection.

[0108] The second stage is kinematic extrapolation. When the duration exceeds the loss-holding duration but does not exceed the sum of the loss-holding duration and the extrapolation duration by 0.5 seconds, the target position is estimated in the world fixed coordinate system by multiplying the target velocity by the duration exceeding the loss-holding duration, and then reprojected back to the current body coordinate system as a temporary target for continued control.

[0109] The third level is to wait in place. When the duration exceeds the sum of the above two, wait in place at zero speed for 0.5 seconds. Its function is to avoid paying the price of rotation for a brief interruption.

[0110] Level 4, directional rotation search. If no valid target is found after the waiting period, rotate in place at the search angular velocity, prioritizing the side where the target last appeared, forcing the forward linear velocity to zero, and accumulating the rotation angle.

[0111] Level 5, Relocking within the Same Category. Only in the Level 4 directional rotation search phase are the marker matching conditions relaxed: any candidate marker whose category matches the locked target and whose world fixed coordinate system position falls within 1.0 meter of the relocking position threshold calculated from the target velocity estimate can be relocked as a following target. After relocking, the smoother state is cleared. In non-search states, matching is still strictly based on the tracking marker to prevent other nearby objects of the same type from preempting the following. When the cumulative rotation angle reaches the maximum search angle, or the total search time reaches the search timeout period, the target is determined to be lost, zero velocity is output, and the current following task ends.

[0112] Step S2, the graded speed shaping and safe zero-speed holding process, involves performing the following operations sequentially on the speed command obtained in step S3: (1) Limit the bandwidth according to the upper limit of each channel; (2) When the forward linear velocity is positive, the turning and deceleration process is performed according to the amplitude of the yaw rate in accordance with formula (7) to reduce the forward linear velocity; , In equation (7), It represents the lower limit of the retention ratio of forward linear velocity during a large turn, and is dimensionless; The angular velocity threshold for deceleration during steering is expressed in radians per second; the arrow indicates that the result on the right replaces the variable on the left. Equation (7) is only used when... When applied, its function is to reduce the forward speed of the robot when making large turns, thereby reducing the risk of tilting of the legged platform under the combined motion of turning and forward movement.

[0113] (3) Calculate the forward safety detection distance according to formula (8), and traverse the cost map grids within the rectangular area in front of the robot that is within this distance and whose half width is half the detection width (skip the grids with target body cost removal marks); if there are occupied grids, or the cost map and coordinate transformation are unavailable, set the forward linear velocity to zero and set the safety stop sign. , In equation (8), The reference speed used to estimate the braking distance is expressed in meters per second. It is the larger of the command speed in the current cycle and the actual speed issued in the previous cycle. Its function is to prevent the detection distance from shrinking prematurely during deceleration. The three terms in equation (8) correspond to the fixed safety margin, the braking distance estimated based on the reaction time, and the braking distance estimated based on the upper limit of deceleration, respectively. The larger of the three is taken.

[0114] (4) Apply the executable lower limit raising to the three channels respectively according to formula (9); , In equation (9), u represents any channel component of forward linear velocity, lateral linear velocity, or yaw angular velocity; This indicates the minimum executable speed amplitude of the channel, the value of which is determined by the minimum speed at which the foot motion control unit can generate an effective step. To represent a symbolic function, the following rules apply. .

[0115] Equation (9) is the core formula of this invention, which has three properties: Firstly, its lifting effect. When hour, Elevated to The sign remains unchanged, allowing the instruction to fall into the executable range of the foot-based platform.

[0116] Secondly, it maintains zero. Because... ,when The result of equation (9) is always zero. This means that the zero velocity generated by the zeroing of the error dead zone and the zero velocity generated by the zeroing of the front safety gate will not be raised into motion commands in this step. Zeroing is the formal basis for this invention to allow the lower limit raising to be placed after the front safety gate; if zeroing is not present, or if the raising is placed before the front safety gate, the channel after the front safety gate is zeroed will be raised again, and the safety judgment will be ineffective.

[0117] Third, idempotency. For those already satisfied... The component of the equation (9) does not change its value, and therefore does not interfere with the output of the control law within the normal speed range.

[0118] In the engineering implementation, a sufficiently small positive number is used (in this embodiment, we take...). The absolute value is used as the threshold for determining zero; values ​​whose absolute value does not exceed this threshold are considered zero.

[0119] (5) Apply acceleration / deceleration limits to the three channels respectively according to formula (10); , In equation (10), This indicates the actual forward linear velocity issued in the previous control cycle, in meters per second. This indicates the control period, in seconds. Lateral linear velocity and yaw rate are limited in the same way, with the upper limits for acceleration and deceleration of lateral linear velocity selected separately depending on whether the velocities increase in the same direction.

[0120] Equation (10) is placed after Equation (9). Its function is to expand the speed step generated by the rise of Equation (9) into a gradual rise within several control cycles, so as to avoid the step command impacting the gait controller. After Equation (10), the channel where the safety stop sign is set must be subjected to a second forced zeroing after the post-set. Its function is to prevent the deceleration upper limit from causing the forward speed to decay to zero within several cycles. The safety stop requirement takes effect immediately and is not subject to the acceleration and deceleration limit constraints.

[0121] (6) If the safety stop sign is in place, the forward linear velocity will be forcibly reset to zero again.

[0122] Step S5: Send the final speed command obtained after the above steps to the leg motion control unit, and report the current status, the identification and type of the locked target, the target distance and azimuth, as well as the forward error, lateral error, angle error and stable frame count under Task B.

[0123] The following describes the typical operating process of this embodiment.

[0124] Scenario 1: Scene X follows the straight line and turns through a right-angle bend.

[0125] The followed object travels straight along the corridor and then turns into a side corridor. When the target distance is greater than the proximity threshold, the robot is in potential field mode: the historical trajectory in the world fixed coordinate system has recorded the complete route before and after the object turns, and the guide point selected according to equation (4) always falls on this route. Therefore, the robot turns along the route taken by the object, rather than tangent to the corner along a straight line. If the target state is maintained in the body coordinate system, then when the robot turns a right angle, the historical trajectory in the body coordinate system also turns a right angle as a whole, and the guide point will point to a non-existent position.

[0126] During the movement, the proportional control law outputs a decreasing forward linear velocity as the target approaches the desired distance. When this velocity falls below... When the interval is reached, it is raised to the level specified in equation (9). After the robot takes a step at an executable speed, the distance error immediately decreases to within the dead zone, as shown in equation (6). If the value is set to zero, equation (9) maintains the zero velocity and the robot stops and stands still. This is the two-state behavior of the present invention: "either clearly stationary or moving at an executable speed".

[0127] Working condition 2: Short-term occlusion of scene Y and switching of tracking markers.

[0128] The object being followed passes behind the pillar. Within 0.8 seconds after the object disappears, the system continues control using the last valid observation; within 0.8 to 1.3 seconds, the target position is calculated based on the target velocity obtained in the world fixed coordinate system according to equation (3) and reprojected back to the current body coordinate system. The robot continues to move forward according to the predicted point, and most obstructions will pass through the window naturally. If the obstruction lasts longer, the robot waits in place for 0.5 seconds and then rotates to search in place towards the side where the object last appeared. When the object reappears, the tracker has assigned it a new identifier, and matching by identifier will fail; at this time, the same category relock condition in the search state of the five-level graded loss recovery subprocess takes effect: the candidate category is consistent with the locked target, and its world fixed coordinate system position falls within 1.0 meter of the predicted position, so it is relocked, the smoother is cleared, and following continues. In the non-search state, this relaxed condition does not take effect, so other similar objects passing by nearby will not take over the following.

[0129] Scenario 3: Orientation of the target on the Z-desktop in the scene.

[0130] The target category assigned in the task belongs to the set of retrievable objects, and the system enters the phase-based pose alignment process. When the robot approaches the tabletop, the target azimuth angle is large, exceeding [a certain value] for three consecutive frames. The moving-in orientation is aligned with the phase and rotated in place; the angular error is reduced to The following is a forward approximation phase, which approximates the target with forward linear velocity, and during this period, small angular velocities with limited lateral error superposition are used for auxiliary correction. When the forward distance is less than the lateral priority threshold and the lateral error exceeds the lateral priority entry threshold, the lateral alignment phase is entered in advance: first, the forward linear velocity is reduced to zero through the zero-velocity transition phase, and then the lateral linear velocity is output according to equation (11), and its value is also raised to the minimum executable amplitude of the lateral channel by equation (9). The orientation of the robot remains unchanged throughout the lateral alignment process, and the relative attitude relationship established between the robot arm and the target is not destroyed. After the lateral error is included in the tolerance, the forward approximation phase is returned to complete the final forward approximation. The forward and lateral errors are included in the tolerance at the same time and are maintained for 10 consecutive frames before the alignment is determined to be complete and reported as ready for capture.

[0131] In this working condition, the target body cost elimination process is ineffective, so the table edge adjacent to the target object will still be identified as an obstacle by the front safety gate; at this time, the forward linear velocity is set to zero and the safety stop sign is set. This sign makes the subsequent second forced zeroing after equation (10) effective, and the robot stops at the specified time without continuing to rush forward due to the deceleration limit. If this state continues for more than the blocking timeout period, the system ends the task and reports a safe stop.

[0132] Condition 4: Local minima of the potential field.

[0133] When the robot moves between two obstacles, the repulsive and attractive forces nearly cancel each other out, and the total potential force in equation (5) approaches zero. The output speed is below the local minimum threshold for a long time, and the target is not within close range. Based on this, the system determines that it has entered a local minimum and enters a timed alternating direction rotation to escape the obstacle for 1.0 seconds. The forward linear velocity is set to zero, and the rotation direction is alternately reversed each time it is triggered to avoid repeatedly hitting the same obstacle. During the escape, the output angular velocity is also increased by equation (9) to ensure that the footed platform actually rotates instead of being stopped and spinning in place.

[0134] The above description is merely illustrative of the embodiments of the present invention and is not intended to limit the present invention. For those skilled in the art, any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A target following control method for a legged humanoid robot, characterized in that, Includes the following steps: Step S1: Obtain the target observation information output by the robot's perception, as well as the robot's body pose in the world fixed coordinate system; Step S2: Perform graded speed shaping and safety zero speed holding processing; first, based on the error dead zone, zero the error smaller than the tolerance is judged, then the executable lower limit is raised for non-zero speeds, then acceleration and deceleration amplitude shaping is performed, and finally, a second forced zeroing is applied after the safety trigger, and the three-degree-of-freedom body speed command executable by the leg motion control unit is output. Step S3: Select the control mode branch based on the target category; if the target is identified as a retrievable object, enter the phased pose alignment process, use the omnidirectional translation capability of the legged robot to eliminate errors by phase division, and perform zero-speed transition before switching between different motion modes; if the target is a dynamic follower, perform dynamic follower control. Step S4: Maintain the target state and historical trajectory in the world fixed coordinate system, and perform target body cost removal processing; When a target is detected as lost, a five-level loss recovery subprocess is triggered to retrieve the target. Step S5: Send the final speed command obtained after the above steps to the foot motion control unit.

2. The target following control method for a legged humanoid robot according to claim 1, characterized in that, In step S2, during the shaping and zero-speed maintenance of the graded speed, the execution lower limit raising has a zero-preservation characteristic; speeds that are greater than zero and less than the minimum execution amplitude of the corresponding channel are raised to the minimum execution amplitude of that channel only. The velocity remains zero when it is zero. After the lifting, acceleration and deceleration amplitude limiting and shaping are performed, and the secondary forced zeroing for safe stopping is set after the acceleration and deceleration amplitude limiting.

3. The target following control method for a legged humanoid robot according to claim 1, characterized in that, The phased pose alignment sub-process described in step S3 includes an orientation alignment phase, a forward approximation phase, a lateral alignment phase, a zero-velocity transition phase, and a stability determination phase. The phase transition uses a hysteresis condition composed of an entry threshold and an exit threshold, and also needs to meet the continuous frame count condition. The lateral alignment phase outputs a lateral translation speed to eliminate lateral errors and keeps the body orientation unchanged throughout the process. When switching motion modes, the zero-velocity transition phase must be entered, and the new mode can only be activated after the speed of the previous mode returns to zero.

4. The target following control method for a legged humanoid robot according to claim 1, characterized in that, The five-level graded loss recovery sub-process described in step S4 is executed sequentially: short-term hold, kinematic extrapolation, in-situ waiting, directional rotation search, and relocking of the same category; the matching conditions are relaxed only when the directional rotation search stage is in progress, allowing relocking of the target based on the target category and the position threshold in the world fixed coordinate system; in the non-search state, the target is strictly matched based on the tracking identifier.

5. The target following control method for a legged humanoid robot according to claim 1, characterized in that, The target cost culling described in step S4 only takes effect in dynamic follow mode; with the target position in the world fixed coordinate system as the center and a preset culling radius, the potential field repulsion force calculation and forward safety gating traversal of the corresponding grid are skipped; the target cost culling is turned off in pose alignment mode.

6. The target following control method for a legged humanoid robot according to claim 1, characterized in that, Between step S1 and step S2, there is also a target observation smoothing sub-step; the target observation smoothing sub-step includes: applying sliding window mean filtering to the forward and lateral coordinates of the target under the machine system; applying first-order low-pass filtering with the shortest angular distance to the target azimuth angle; applying first-order low-pass filtering to the target distance to suppress periodic observation jitter caused by gait fluctuations.

7. The target following control method for a legged humanoid robot according to claim 1, characterized in that, The dynamic following control described in step S3 is as follows: When in dynamic following mode, guide point potential field obstacle avoidance control is adopted; guide points are selected based on the target's historical trajectory to generate attraction, and obstacle grids generate repulsion and tangential bypass forces; angle adaptive synthesis weights are configured, and the greater the target's azimuth angle deviates from directly in front, the higher the attraction weight; it also includes potential field local minima detection, and when the local minima is determined, time-limited alternating direction rotation in place is performed to escape from the obstacle.

8. A target following control method for a legged humanoid robot according to claims 1-7, characterized in that, Step S2 also includes steering deceleration processing, specifically: when the forward linear velocity is greater than zero, the forward linear velocity is reduced according to the yaw rate amplitude ratio; the detection distance of the forward safety gate is the maximum value among the fixed safety distance, the reaction braking distance, and the dynamic estimated braking distance.

9. A target following control system for a legged humanoid robot according to claim 1, used to implement the control method according to any one of claims 1-8, characterized in that, include: The visual perception component is used to output the 3D bounding box, category label, confidence score, and tracking marker of the target within the field of view; Environmental perception components are used to construct local cost maps; The robot pose estimation component is used to output the robot's pose and velocity in a fixed world coordinate system. The foot motion control unit is used to receive three-degree-of-freedom body velocity commands and calculate the foot landing point and joint trajectory; An embedded computing unit is used to run the follow control program; the embedded computing unit internally deploys the following modules: The target candidate filtering module is used to filter out target candidates that do not meet the following conditions from the perceived output targets; The stable locking and jump rejection module is used to achieve stable target locking and reject abnormal jump observations; The target observation smoothing module is used to filter target observation data and suppress periodic jitter caused by gait fluctuations. The World Fixed Coordinate System Target State Estimation Module is used to calculate the target position and velocity in the world fixed coordinate system. The path memory and guide point module is used to record the target's historical trajectory and generate follow guide points; The target cost elimination module is used to remove the grid corresponding to the target being followed from the cost map repulsion force and safety judgment. The control mode arbitration module is used to select the corresponding control law based on the distance between the task and the target. The close-range proportional-integral control module is used to calculate the following speed command under close-range operating conditions; The phased pose alignment module is used to execute the phased pose alignment sub-process for the target object; The potential field obstacle avoidance control module is used to implement potential field obstacle avoidance following control based on the guide point and cost map; The speed limiting and steering speed reduction module is used to limit the control output speed and reduce the forward speed during sharp turns. The forward safety gating module is used to detect obstacles in front of the robot, trigger the forward speed to zero and set the safety stop sign; An executable lower limit lifting module is used to perform zero-preservation executable lower limit lifting processing on speed commands. The acceleration / deceleration limiting and safety zeroing module is used to perform acceleration / deceleration limiting and to force zeroing a second time after safety triggering. The graded loss recovery module is used to execute the five-level graded loss recovery sub-process to complete the target retrieval and disposal; The health gating and emergency stop module is used to monitor the health status of the system and output zero speed when an emergency stop or posture timeout occurs. The task session and status reporting module provides interfaces for starting and stopping tasks, and reports task running status and target information.

10. A target following control system for a legged humanoid robot according to claim 9, characterized in that, The visual perception component outputs to the target candidate filtering module; the target candidate filtering module outputs to the stable locking and jump rejection module; the stable locking and jump rejection module outputs to the target observation smoothing module and the hierarchical loss recovery module; The target observation smoothing module outputs to the world fixed frame target state estimation module; the body pose estimation component outputs to the world fixed frame target state estimation module; the world fixed frame target state estimation module outputs to the path memory and guide point module, the target body cost elimination module, and the hierarchical loss recovery module respectively; the environmental perception component outputs to the target body cost elimination module and the potential field obstacle avoidance control module; the control mode arbitration module receives outputs from the path memory and guide point module, the task session and status reporting module, and the hierarchical loss recovery module, and outputs to the near-range proportional integral control module. The control module, the phased pose alignment module, and the potential field obstacle avoidance control module; the outputs of the near-range proportional-integral control module, the phased pose alignment module, and the potential field obstacle avoidance control module are sent to the speed limiting and steering deceleration module; the output of the speed limiting and steering deceleration module is sent to the forward safety gating module; the output of the forward safety gating module is sent to the executable lower limit raising module; the output of the executable lower limit raising module is sent to the acceleration / deceleration limiting and safety zeroing module; the output of the health gating and emergency stop module is sent to the acceleration / deceleration limiting and safety zeroing module; and the output of the acceleration / deceleration limiting and safety zeroing module is sent to the foot motion control unit.

Citation Information

Patent Citations

  • Robot following obstacle avoidance method and device and storage medium

    CN115509231A

  • Robot following obstacle avoidance method and related equipment

    CN117193327A

  • Robot following obstacle avoidance method and related equipment

    CN117420828A

  • Method for humanoid robot to follow accompanying object

    CN120686839A

  • Composite robot target following method and related equipment

    CN121552396A