Humanoid robot tension control method and device based on improved contact model

Abstract

The invention belongs to the field of humanoid robots, and discloses a humanoid robot tension control method and device based on an improved contact model. According to the method, a four-layer hierarchical control framework is constructed, an improved contact force model is provided, a smooth continuous function is adopted to replace traditional rigid complementary constraint, the first item is sigmoid-based repulsive force, the second item is Gaussian attractive force, and tension (negative normal force) is allowed to be generated at a contact point. An upper layer heuristic gait planner rapidly generates a foot falling position through dead zone feedback; the middle-layer CIMPC combines and optimizes a mass center track and a contact force sequence through a simplified single rigid body model; mapping the whole body controller into joint instructions according to task priorities; and the PD servo controller at the bottom layer drives the joints at high frequency. According to the method, the humanoid robot can walk in a complex environment such as a low-friction slope and a narrow channel in an auxiliary mode through adsorption tension, and the movement capacity and robustness are remarkably improved.

Description

Technical Field

[0001] This invention belongs to, but is not limited to, the field of humanoid robot technology, and particularly relates to a method and device for controlling the pulling force of a humanoid robot. Background Technology

[0002] With the development of humanoid robot technology, robots are gradually demonstrating broad application prospects in fields such as home service, industrial collaboration, and disaster relief. As a system with high degrees of freedom and flexible adaptability, humanoid robots can complete diverse tasks in complex environments. However, due to the dynamic characteristics of humanoid robots themselves, their complex kinematic models, and environmental interaction constraints, existing motion control methods have some limitations in practical applications. Especially when dealing with nonlinear dynamics, contact forces, and environmental constraints, existing control methods struggle to effectively meet the motion requirements of highly complex environments.

[0003] To address these issues, researchers have proposed various simplified models and control frameworks, such as zero-moment point (ZMP) gait planning based on the linear inverted pendulum (LIP) model, hybrid zero dynamics (HZD), and hybrid inverted pendulum (HLIP) regulators. These simplified models transform complex nonlinear problems into easily solvable linear problems through virtual constraints or linear approximations, thereby generating feasible gait trajectories. However, these simplified models exhibit significant dynamic differences from real robots, and the generated trajectories cannot be guaranteed to remain stable on real robots. Especially for walking and interaction in highly dynamic and complex environments, traditional methods suffer from poor efficiency and robustness.

[0004] To further optimize control methods, Model Predictive Control (MPC) has been widely applied to legged robots in recent years. MPC can solve for optimal control sequences in real time within a limited prediction window, responding to environmental disturbances and dynamic constraints. Contact-Implicit MPC further extends LIP-MPC to contact-rich environments, describing contact dynamics through bi-level programming to optimize motion trajectories and contact timing. This method can generate new contact patterns in real time and exhibits good robustness, making it widely used in quadrupedal and bipedal robot systems.

[0005] At the control level, Whole-Body Control (WBC) coordinates multiple tasks using task prioritization and optimization methods to achieve balance and coordination in actions such as walking, grasping, and supporting. Lower-level control typically uses proportional-derivative (PD) controllers or impedance controllers to drive robot joints, ensuring the efficiency and stability of high-frequency actuation. For movement in complex environments, robots can also utilize negative pressure devices such as suction cups to provide additional adhesion.

[0006] Humanoid robots, as highly flexible and adaptable robots, have demonstrated enormous potential in multiple fields. However, due to their complexity, especially in motion control in complex environments, humanoid robots still face several technical challenges. Traditional motion control methods largely rely on simplified models, such as gait planning methods based on the inverted pendulum model and contact implicit model predictive control (CIMPC). Although these methods have made progress in stability and computational efficiency, their ability to handle humanoid robots in complex environments remains insufficient, particularly in handling contact forces, the real-time performance of gait planning, and the efficiency of task coordination.

[0007] To adapt to different environmental changes and task requirements, improvements to contact models, optimization of gait planning methods, and the introduction of task scheduling mechanisms have become important research directions in the field of humanoid robot control. The traditional CIMPC model only allows normal forces when handling interactions at contact points. This limitation is particularly prominent in scenarios where robots need to interact strongly with their environment, such as walking along walls, ramps, or low-friction surfaces. Furthermore, the real-time performance and computational burden of gait planning are also exposed in highly dynamic environments, affecting the robot's ability to quickly adapt to environmental changes.

[0008] Against this backdrop, this invention proposes improvements to existing technologies, particularly in gait planning, contact models, and task coordination, by optimizing the control architecture to address the shortcomings of existing technologies:

[0009] (1) The contact model is limited by the inability to generate tension: The traditional CIMPC contact model only allows the generation of normal force (i.e., repulsive force) at the time of contact. The contact force function increases monotonically as the contact distance approaches zero, and there is no attraction interval. This means that the robot can only maintain stable contact with the environment through thrust. In scenarios where the robot needs to rely on the suction force of its hands to walk stably (such as climbing low-friction slopes or walking while holding onto walls), the traditional model cannot generate tension, resulting in insufficient system stability and adaptability.

[0010] (2) Lack of intermediate mapping layer for task coordination: Existing architectures typically generate joint trajectories directly using full-body MPC and track them through the underlying servo system. The lack of an independent full-body control layer results in poor task coordination when handling multiple tasks. Since MPC must directly process all joint variables, it cannot fully utilize the task priority mechanism, which limits the system's flexibility. Especially when multiple tasks are involved or when interacting with the environment, it cannot effectively adjust the priorities between tasks, leading to insufficient system coordination.

[0011] (3) Insufficient real-time performance of gait planning: In existing technologies, gait planning usually relies on an HLIP regulator combined with a complete CIMPC model. The HLIP regulator needs to solve an optimization problem in each gait cycle, so the response speed is slow under rapidly changing speeds or pushing disturbances, especially when facing complex environments, the real-time performance of gait planning cannot meet the requirements. The limitation of this method is that it cannot quickly adapt to changes in the environment or the dynamic changes in robot movements.

[0012] (4) Excessive computational burden: The traditional CIMPC model requires simultaneous optimization of all joint states and contact forces, which is complex and involves a large number of parameters and computations. Due to the need to accurately calculate each joint and contact force, the computational burden is heavy and it is easy to get trapped in local extrema. In addition, the optimization of the traditional model requires a lot of parameter tuning to generate a reasonable gait, which affects the real-time performance and ease of adjustment of the system and makes it unable to react quickly in highly dynamic environments.

[0013] Based on the above analysis, the urgent technical problems that need to be solved in the existing technology are:

[0014] Existing technologies cannot simultaneously meet the requirements of real-time performance, computational efficiency, contact diversity, and task coordination, making it difficult to satisfy the motion capabilities required of humanoid robots in complex three-dimensional environments. Summary of the Invention

[0015] Terminology Explanation:

[0016]

[0017] To address the problems existing in the prior art, the present invention provides a method and apparatus for controlling the pulling force of a humanoid robot.

[0018] This invention is implemented as follows: a humanoid robot tension control method based on an improved contact model, specifically including:

[0019] Step 1: Collect the robot's center of mass state and contact state. The center of mass state includes position, velocity, and orientation. The contact state includes the signed distance between each contact point. and adsorption pressure;

[0020] Step 2: Based on the dynamics of the linear inverted pendulum (LIP) and the dead zone feedback mechanism, generate the landing target and the hand contact target according to the center of mass state and the desired motion requirements;

[0021] Step 3: Within the prediction window, simplify the robot dynamics to a single rigid body model, and use the generalized coordinate trajectory q(0:N), generalized velocity trajectory v(0:N), and joint torque tau(0:N-1) as optimization variables. Under the conditions of satisfying dynamic constraints and friction constraints, jointly optimize the center of mass trajectory and contact force sequence.

[0022] Step 4, in the prediction process, the contact force is calculated using an improved smooth continuity function:

[0023] The first term is a sigmoid-based repulsive force term, and the second term is a Gaussian attractive force term; when the contact distance... When ≈0, the attraction term generates a pulling force, simulating the suction cup adsorption effect; when the contact point is a normal foot support, setting ka=0, the contact force function degenerates into a traditional model that only repels.

[0024] Step 5: Based on the optimized centroid trajectory and contact force, map them into whole-body joint targets and adsorption pull targets according to task priority;

[0025] Step 6: Send the joint target and the adsorption pull target to the execution layer to drive the robot's actions; feed back the actual state to the prediction layer to form a closed-loop update.

[0026] Furthermore, the calculation of the landing target is based on the feedforward-feedback structure of the linear inverted pendulum (LIP) model, and the step size calculation formula is as follows: ,in Let T be the desired walking speed and T be the gait period. and These represent the position and velocity of the robot's center of mass at the moment of step change. and This is the corresponding reference value for the LIP model. and This is the dead-zone feedback gain matrix. The swing foot trajectory is generated using 7th-order Bezier curve interpolation.

[0027] Furthermore, the improved contact force function c( The values ​​for each parameter in the formula are as follows: repulsion stiffness kr ranges from 500 to 5000 N / m, sigmoid sharpness... The value ranges from 5 to 20, and the attraction intensity ka ranges from 0.1 to 10 N, representing the attraction range. The values ​​range from 0.001 to 0.01 m; in the CIMPC solver, the contact stiffness parameter ranges from 2000 to 5000 N / m, the dissipation velocity ranges from 0.05 to 0.2 m / s, the smoothing factor ranges from 0.005 to 0.01, and the friction coefficient... The value ranges from 0.5 to 0.7; the friction cone constraint is ||ft|| <= * max(c( ), 0).

[0028] Another object of the present invention is to provide a humanoid robot pulling force control device, comprising:

[0029] The state sensing unit is used to acquire the state of the center of mass and the contact state;

[0030] The predictive optimization unit is used to jointly optimize the centroid trajectory and contact force based on a single rigid body dynamics model and a tensile contact model that allows negative normal forces.

[0031] The task mapping unit is used to convert the center of mass trajectory and contact force into joint motion targets and adsorption pull targets according to task priority.

[0032] An execution control unit is used to drive the robot's movements according to the joint motion target and the adsorption pull target;

[0033] The feedback unit is used to feed back the execution status to the prediction and optimization unit to form a closed loop.

[0034] In a preferred embodiment, a negative pressure adsorption device is installed at the end of a robot hand and includes a flexible sealed cavity, a negative pressure generating unit, and a pressure sensor to generate a controllable adsorption pull at the contact surface; the negative pressure generating unit may be an electrically driven vortex fan or a micro vacuum pump.

[0035] Furthermore, the prediction and optimization unit simultaneously applies positional constraints and contact friction constraints between the centroid and the support region.

[0036] Furthermore, the task mapping unit adopts a hierarchical priority constraint method, prioritizing the execution of balanced tasks, followed by landing tasks and contact tasks.

[0037] Another object of the present invention is to provide a whole-body control mapping method for tension control of a humanoid robot, comprising:

[0038] Receive the centroid trajectory and contact force generated by the prediction optimization unit;

[0039] The centroid trajectory is converted into a balance constraint in the joint space;

[0040] The contact force is converted into force targets at the hand and foot ends;

[0041] According to the preset task priority order, the balance constraint is satisfied first, then the landing position constraint and contact posture constraint are satisfied, and finally the adsorption pull constraint is satisfied.

[0042] Output the joint angle target, joint torque target, and adsorption pull target that satisfy the aforementioned priority constraint relationship.

[0043] Furthermore, the adsorption pull target is dynamically adjusted based on the deviation between the predicted contact force and the actual adsorption pressure.

[0044] Furthermore, the joint angle target and the joint torque target are solved simultaneously through a constraint optimization method.

[0045] Furthermore, the task priority is dynamically adjusted when the risk of robot instability increases, so that the priority of balancing tasks is higher than that of contact tasks and motion tasks.

[0046] The following is a comparative analysis of this invention with the closest prior art:

[0047]

[0048] Based on the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by this invention are as follows:

[0049] First, this invention addresses the shortcomings of existing technologies by proposing a novel robust motion control architecture, comprising an upper-layer heuristic gait planning layer, a middle-layer simplified CIMPC combined with a complete WBC model, and a lower-layer improved tension-contact model and PD controller. The advantages of each improvement and the solutions it addresses are analyzed below:

[0050] (1) Achieved contact modeling with tension: The traditional CIMPC contact force model only supports non-negative normal forces (repulsive forces). This invention proposes a smooth and continuous repulsive-attraction composite contact force function, which is achieved through the Gaussian attraction term. The adhesion pull is naturally generated at approximately 0, allowing the optimizer to automatically utilize the adhesion pull during the planning process. This function is continuously differentiable, avoiding the discontinuities in complementary constraints during state transitions and thus contributing to the numerical stability of the optimization process.

[0051] (2) Improved real-time performance of gait planning: In existing technologies, traditional gait planning often relies on a combination of HLIP regulators and CIMPC models. HLIP regulators need to solve complex nonlinear optimization problems in each gait cycle, which is computationally intensive and results in slow response under rapid speed changes or pushing disturbances. This invention adopts a dead-zone feedback heuristic gait planner based on LIP dynamics. It quickly generates the foot position and nominal gait by analytical step length correction based on the current centroid state and desired speed. It does not require online solving of the complex optimization problems in existing gait regulators, reducing the response time by an order of magnitude and significantly improving the real-time performance of gait planning.

[0052] (3) Reduced complexity of optimization calculation: In the existing technology, CIMPC uses a complete multi-rigid-body model, which requires simultaneous optimization of the state and contact force of all joints of the robot. The calculation process is huge, involves a large number of parameters, and is prone to getting trapped in local extrema. The present invention simplifies the robot into a single rigid-body model (27-dimensional generalized coordinates and 26-dimensional generalized velocity), which greatly reduces the number of optimization variables. With the hot start and two iteration upper limit, the single-step MPC solution time is controlled in the millisecond level, which meets the 50Hz real-time execution requirement.

[0053] (4) Enhanced multi-task coordination capability: In existing technologies, the CIMPC control architecture typically generates joint trajectories directly from the whole-body MPC. The MPC must process all joint variables, resulting in low computational efficiency and potential conflicts when handling multiple tasks such as balance, foot position, arm posture, and suction. This invention introduces an independent WBC layer and uses task priority QP solution, prioritizing center of mass balance > foot contact > hand contact > posture maintenance. This allows for the completion of low-priority tasks as much as possible without affecting high-priority tasks, significantly improving the system's task coordination capability.

[0054] In summary, the hierarchical control architecture proposed in this invention improves upon the shortcomings of existing technologies one by one, providing well-founded optimizations from gait planning, predictive control, contact modeling to task coordination and execution layer control, thereby significantly enhancing the mobility and robustness of humanoid robots in complex environments.

[0055] Secondly, as supporting evidence of the inventiveness of this invention, it is also reflected in the following important aspects:

[0056] (1) The expected benefits and commercial value of the technical solution of the present invention after transformation are as follows: The present invention can be applied to scenarios such as building construction, fire rescue, and industrial inspection where robots need to climb in complex terrain. By enabling humanoid robots to use adsorption and pulling force to assist walking, the operating range of robots can be expanded and the risk of human exposure in dangerous environments can be reduced.

[0057] (2) The technical solution of this invention fills a technical gap in the industry at home and abroad: the contact force model of the existing contact implicit model predictive control framework is a pure repulsion model (contact force is non-negative), and the attraction mechanism is not included in the optimization framework. This invention introduces a Gaussian attraction term in CIMPC for the first time, and constructs a smooth and continuous repulsion-attraction composite contact force function, so that the optimizer can automatically utilize the adsorption force, filling the technical gap of tensile force modeling in contact implicit MPC.

[0058] (3) The technical solution of the present invention solves the technical problem that people have long wanted to solve but have never been able to solve: In scenarios such as low-friction slopes, relying solely on foot friction is insufficient to maintain the stable walking of humanoid robots. The present invention solves the problem of coordinated distribution of hand and foot contact forces by incorporating attraction into the contact force function of CIMPC, enabling the optimizer to automatically allocate the optimal ratio of foot repulsion force and hand adsorption force when planning the centroid trajectory.

[0059] (4) The technical solution of the present invention overcomes technical bias: In the prior art, it is generally believed that the contact force model must ensure that the normal force is non-negative (only repulsive). The present invention overcomes this bias by ensuring the boundedness of the Gaussian attraction term (peak value). Limited scope of application The finite (limited) guarantee of the well-defined nature of the optimization problem proves that introducing attractive forces under appropriate parameters will not lead to optimization divergence. Attached Figure Description

[0060] Figure 1 This is a flowchart of a humanoid robot tension control method based on an improved contact model provided in an embodiment of the present invention;

[0061] Figure 2 This is a four-layer hierarchical control architecture diagram provided in an embodiment of the present invention, which includes a gait planner (50Hz), CIMPC (50Hz), WBC (500Hz) and PD servo (500Hz), with the inputs and outputs of each layer labeled;

[0062] Figure 3 This is an improved contact force function curve provided in an embodiment of the present invention;

[0063] Figure 4 This is a diagram illustrating walking on a slope with support against a wall and force analysis provided in an embodiment of the present invention;

[0064] Figure 5 This is the LIP dead zone feedback heuristic step size calculation diagram provided in the embodiments of the present invention;

[0065] Figure 6 This is a 7th-order Bezier swing foot trajectory diagram provided in an embodiment of the present invention. Detailed Implementation

[0066] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0067] like Figure 1 As shown in the figure, an embodiment of the present invention provides a humanoid robot pulling force control method based on an improved contact model, which specifically includes:

[0068] S101, Collect the robot's center of mass state and contact state. The center of mass state includes position, velocity, and attitude. The contact state includes the signed distance between each contact point. and adsorption pressure;

[0069] S102, based on the dynamics of the linear inverted pendulum (LIP) and the dead zone feedback mechanism, generates the landing target and the hand contact target according to the center of mass state and the desired motion requirements;

[0070] S103, within the prediction window, the robot dynamics are simplified to a single rigid body model, and the generalized coordinate trajectory q(0:N), generalized velocity trajectory v(0:N), and joint torque τ(0:N-1) are used as optimization variables. Under the conditions of satisfying dynamic constraints and friction constraints, the center of mass trajectory and contact force sequence are jointly optimized.

[0071] S104, in the prediction process, the contact force is calculated using an improved smooth continuity function:

[0072] The first term is a sigmoid-based repulsive force term, and the second term is a Gaussian attractive force term; when the contact distance... When ≈0, the attraction term generates a pulling force, simulating the suction cup adsorption effect; when the contact point is a normal foot support, setting ka=0, the contact force function degenerates into a traditional model that only repels.

[0073] S105, Based on the optimized centroid trajectory and contact force, it is mapped into whole-body joint targets and adsorption pull targets according to task priority;

[0074] S106, the joint target and the adsorption pull target are sent to the execution layer to drive the robot's actions; the actual state is fed back to the prediction layer to form a closed-loop update.

[0075] The target landing calculation provided in this embodiment of the invention is based on the feedforward-feedback structure of the linear inverted pendulum (LIP) model, and the step size calculation formula is as follows: ,in Let T be the desired walking speed and T be the gait period. and These represent the position and velocity of the robot's center of mass at the moment of step change. and This is the corresponding reference value for the LIP model. and This is the dead-zone feedback gain matrix. The swing foot trajectory is generated using 7th-order Bezier curve interpolation.

[0076] The improved contact force function c( ) provided in this embodiment of the invention The values ​​for each parameter in the formula are as follows: repulsion stiffness kr ranges from 500 to 5000 N / m, sigmoid sharpness... The value ranges from 5 to 20, and the attraction intensity ka ranges from 0.1 to 10 N, representing the attraction range. The values ​​range from 0.001 to 0.01 m; in the CIMPC solver, the contact stiffness parameter ranges from 2000 to 5000 N / m, the dissipation velocity ranges from 0.05 to 0.2 m / s, the smoothing factor ranges from 0.005 to 0.01, and the friction coefficient... The value ranges from 0.5 to 0.7; the friction cone constraint is ||ft|| <= * max(c( ), 0).

[0077] An embodiment of the present invention provides a humanoid robot pulling force control device, comprising:

[0078] The state sensing unit is used to acquire the state of the center of mass and the contact state;

[0079] The predictive optimization unit is used to jointly optimize the centroid trajectory and contact force based on a single rigid body dynamics model and a tensile contact model that allows negative normal forces.

[0080] The task mapping unit is used to convert the center of mass trajectory and contact force into joint motion targets and adsorption pull targets according to task priority.

[0081] An execution control unit is used to drive the robot's movements according to the joint motion target and the adsorption pull target;

[0082] The feedback unit is used to feed back the execution status to the prediction and optimization unit to form a closed loop.

[0083] In a preferred embodiment, a negative pressure adsorption device is installed at the end of a robot hand and includes a flexible sealed cavity, a negative pressure generating unit, and a pressure sensor to generate a controllable adsorption pull at the contact surface; the negative pressure generating unit may be an electrically driven vortex fan or a micro vacuum pump.

[0084] The prediction optimization unit provided in this embodiment of the invention simultaneously applies positional constraints and contact friction constraints between the centroid and the support region.

[0085] The task mapping unit provided in this embodiment of the invention adopts a hierarchical priority constraint method, prioritizing the execution of balanced tasks, followed by landing tasks and contact tasks.

[0086] This invention provides a whole-body control mapping method for tension control of a humanoid robot, comprising:

[0087] Receive the centroid trajectory and contact force generated by the prediction optimization unit;

[0088] The centroid trajectory is converted into a balance constraint in the joint space;

[0089] The contact force is converted into force targets at the hand and foot ends;

[0090] According to the preset task priority order, the balance constraint is satisfied first, then the landing position constraint and contact posture constraint are satisfied, and finally the adsorption pull constraint is satisfied.

[0091] Output the joint angle target, joint torque target, and adsorption pull target that satisfy the aforementioned priority constraint relationship.

[0092] The adsorption pull target provided in this embodiment of the invention is dynamically adjusted based on the deviation between the predicted contact force and the actual adsorption pressure.

[0093] The joint angle target and joint torque target provided in this embodiment of the invention are solved simultaneously through a constraint optimization method.

[0094] The task priority provided in this embodiment of the invention is dynamically adjusted when the risk of robot instability increases, so that the priority of balancing tasks is higher than that of contact tasks and motion tasks.

[0095] This invention provides a method for controlling the pulling force of a humanoid robot, the method specifically including:

[0096] S1: Real-time acquisition of the robot's current state information, including base posture and velocity, center of mass motion state, drive joint state, and actuator state information when using attached actuators;

[0097] The sensor data is processed using a 13-dimensional extended Kalman filter (EKF) for state estimation. The state vector includes centroid position (3D), centroid velocity (3D), attitude quaternion (4D), and gyroscope bias (3D). The prediction step uses IMU gyroscope measurements, and the update step fuses accelerometer and gyroscope data. EKF parameters: position process noise 0.001, velocity process noise 0.01, quaternion process noise 0.0001, angular velocity process noise 0.01; acceleration measurement noise 0.5, angular velocity measurement noise 0.05. The filter is automatically reset when the innovation exceeds the threshold (10.0) to prevent filter divergence.

[0098] S2: The heuristic gait planner calculates the next foot placement position based on the current center of mass state, desired velocity, and environmental constraints using an analytical step size correction method, and generates the arm contact target when needed;

[0099] S3: Simplified CIMPC simplifies the robot to a single rigid body model (generalized coordinates nq=27, generalized velocity nv=26). Within a prediction window of N=25 steps and dt=0.05s (prediction time domain 1.25s), CIMPC optimizes the robot's future trajectory and control inputs. Guided by the landing target, hand contact target, and nominal reference provided by the upper-level heuristic gait planner, it combines an improved contact model, dynamic constraints, and friction constraints to jointly optimize the future motion trajectory and contact action. The contact force is calculated using an improved smooth continuous function, allowing for... An attractive force (negative normal force) is generated at approximately 0, simulating the suction cup effect. The solution employs an iterative method based on the trust region, with a maximum of two iterations per cycle, utilizing a warm start to accelerate convergence.

[0100] S4: The whole-body controller solves for joint angles and torque feedforward based on the prediction results and task priorities, and determines the target of the attached actuator;

[0101] S5: The underlying servo controller drives the joint at a high frequency; it uses PD tracking and superimposed torque feedforward to update the hand contact control quantity; when using negative pressure adsorption hardware, the attachment actuator target can be further mapped to the negative pressure generating unit control quantity;

[0102] S6: Feedback loop, which feeds back the actual state to the upper layer and enters the next control cycle.

[0103] This invention, through such a layered design and improved contact modeling, enables humanoid robots to simultaneously achieve rapid gait adjustment, accurate dynamic prediction, stable adsorption and pulling force generation, and whole-body coordinated control in complex environments, meeting the needs of walking and climbing in complex terrains.

[0104] like Figure 2 As shown in the figure, an embodiment of the present invention provides a humanoid robot pulling force control device. The device is divided into four main levels from top to bottom. Each level undertakes a specific task and cooperates with each other, namely:

[0105] Heuristic Gait Planner: Located at the highest level, this planner determines the next foot placement and hand contact target based on the current center of mass state and desired motion requirements. It employs a simple linear calculation algorithm, calculating the next step length correction based on the difference between the robot's current center of mass velocity and the desired velocity. This avoids the drawbacks of traditional HLIP regulators, which require solving complex optimizations in each gait cycle, while also enabling rapid response to speed changes and external disturbances.

[0106] Simplified CIMPC Layer: This layer simplifies the robot to a single rigid body model (generalized coordinates nq=27, generalized velocity nv=26), primarily optimizing the center-of-mass momentum and forces at each contact point to predict the robot's trajectory and contact timing within a future time window (N=25 steps × dt=0.05s=1.25s). This layer embeds an improved contact force function, allowing for adjustments at contact distances... When the radius is approximately 0, an attractive force (negative normal force) is generated. During the prediction process, this layer provides dynamic constraints and friction cone constraints for optimization. The solver uses the trust region iteration method (initial radius 0.1), with a maximum of 2 iterations per cycle, and utilizes hot start to accelerate convergence.

[0107] The Whole Body Controller (WBC) layer is responsible for translating the center-of-mass motion and contact forces generated by the upper-level CIMPC layer into specific joint targets. This layer utilizes a task prioritization strategy, first satisfying the basic task of maintaining balance, and then executing secondary tasks such as foot placement, arm posture, and adhesion. Through this hierarchical mapping, the robot can maintain overall stability while simultaneously performing complex behaviors such as hand-foot coordination.

[0108] The underlying PD servo control layer consists of proportional-derivative (PD) or impedance controllers for each joint, driving the motor at a frequency of 500Hz. The underlying controller adjusts the movement of each joint in real time according to the joint angle and torque commands calculated by the WBC, while simultaneously adjusting the negative pressure of the suction cup via pressure sensors to ensure that the contact force meets the optimized requirements.

[0109] These layers are interconnected through a standard interface, enabling the robot to adaptively adjust its gait, generate dynamically consistent motion trajectories, coordinate whole-body movements, and execute them rapidly in complex environments.

[0110] The goal of the gait planner is to quickly generate the robot's next foot placement, thereby rapidly adapting to external disturbances and speed changes during walking. Its specific implementation steps are as follows:

[0111] (1) State perception: Real-time acquisition of the linear velocity and angular velocity of the robot's center of mass, as well as the current state of the supporting leg and swing leg.

[0112] (2) Step length correction calculation: Based on the difference between the current center of mass velocity and the expected velocity, the step length correction amount is calculated by a linear function, and the next foot landing position is determined accordingly.

[0113] (3) Arm target generation: When the robot needs to contact the environment, such as when it is holding onto a wall or grabbing a bar, the gait planner generates the desired trajectory of the arm based on the position of the center of mass and the geometric information of the environment, so that the hand can contact the target at the right time and control the contact speed, which facilitates the generation of stable pulling force later.

[0114] (4) Output interface: Send the desired foot position, arm target, and reference centroid motion information to the next layer CIMPC. Since this method is based on simple linear functions for calculation, it avoids the dependence on nonlinear optimization in the traditional HLIP method, which greatly improves the response speed and computational efficiency of gait planning.

[0115] The simplified CIMPC layer is responsible for generating a dynamically consistent center-of-mass trajectory and contact force within a certain time prediction window, based on reference information provided by the gait planner. Its core implementation steps include:

[0116] (1) Dynamic modeling: The robot is treated as a single rigid body, considering only the linear and angular momentum of the center of mass, and the contact point position and force are introduced as decision variables. This modeling significantly reduces the number of optimization variables and improves the solution speed.

[0117] (2) Constraint setting: During the optimization process, dynamic balance constraints, position constraints of the center of mass and the support area, friction cone constraints and contact complementary conditions are set to ensure that the generated motion trajectory and contact force meet the physical requirements.

[0118] (3) Improved contact model: Traditional contact models only allow positive normal forces and cannot simulate adhesion forces. This invention adopts a smooth and continuous repulsion-attraction composite contact force function, which includes the signed contact distance. Mapped to normal force. The first term is repulsive force: when When the repulsive force is less than 0 (penetration), it is approximately equal to... *| |Linear growth, when When the value is >0 (separation), the sigmoid decays to zero. The second term is a Gaussian attraction: in Peak values ​​are generated near 0. The tensile force, the range of action is from Control. When the robot's hand engages with the negative pressure suction cup, it causes... >0 causes tension at the contact point; foot contact point causes =0 degenerates into a traditional repulsion model. This function is continuously differentiable, avoiding the discontinuities in traditional complementary constraints during contact state transitions. The default parameters in the code implementation are: =200 N / m, =20, =5 N, =5 mm.

[0119] (4) Optimization solution: Use numerical optimization algorithm to solve the centroid trajectory and contact force in the future view window, so that they are as close as possible to the reference values ​​provided by the gait planner, and ensure that the contact conditions and friction conditions are not violated.

[0120] By simplifying the dynamic modeling and improving the contact model, this invention solves the problem that traditional models cannot generate tensile force while maintaining the advantages of predictive control.

[0121] The centroid trajectory and contact forces generated by the CIMPC need to be mapped to the joints of the entire robot. The specific implementation of the Whole Body Controller (WBC) includes:

[0122] (1) Task decomposition: Based on the robot's task requirements, the control target is divided into multiple tasks, such as center of mass balance, foot position tracking, arm posture adjustment and adsorption force control.

[0123] (2) Priority setting: Assign priority to each task. Usually, the balancing task is the highest priority, followed by the foot and arm tasks, and the suction adjustment task is placed in a lower priority.

[0124] (3) Optimize the solution: Solve the joint angles and torques under the given priority constraints so that the high priority tasks are fully satisfied, and at the same time, complete the low priority tasks as much as possible while satisfying the high priority tasks.

[0125] (4) Output interface: Generate joint angle reference and joint torque command, and pass the target negative pressure value for suction cup to the underlying servo layer.

[0126] Through a task priority mechanism, WBC coordinates multiple tasks simultaneously while ensuring robot balance, thus compensating for the shortcomings of simplified CIMPC, which neglects the details of whole-body motion.

[0127] The underlying PD controller is responsible for converting higher-level commands into actual motor driving force and performing closed-loop control at high frequencies.

[0128] (1) Joint drive: A proportional-differential (PD) or impedance controller is used to calculate the joint error and output the driving force at a sampling rate of 500Hz to ensure that each joint tracks the upper layer command.

[0129] (2) Negative pressure adjustment: The pressure difference of the suction cup cavity is monitored in real time by the pressure sensor, and the output of the negative pressure generator is adjusted according to the target suction force provided by WBC so that the adsorption force meets the optimization requirements.

[0130] (3) Feedback fusion: The feedback from the joint sensor and pressure sensor is sent back to the upper layer so that the gait planner and CIMPC can update the state estimate in the next cycle.

[0131] The underlying servo control is executed at a high frequency to ensure that the robot can smoothly and accurately track the trajectory and force commands generated by the upper layer.

[0132] like Figure 3 Improved contact force function curve;

[0133] like Figure 4 Walking on a slope with support against a wall and force analysis;

[0134] like Figure 5 LIP dead zone feedback heuristic step size calculation;

[0135] like Figure 6 , 7th order Bezier swing foot trajectory.

[0136] Example 1: Stability control for walking with support on a flat indoor floor

[0137] In the simulation verification scenario, a Kuavo S46 humanoid robot model was used. The robot's generalized coordinate dimension nq=27 (4-dimensional base posture quaternion, 3-dimensional base position, 6-dimensional left and right legs, and 4-dimensional left and right arms), and generalized velocity dimension nv=26. The expected walking speed is... =0.25 m / s.

[0138] CIMPC parameter configuration: prediction steps N=25, time step dt=0.05s (corresponding to prediction time domain Tp=1.25s), MPC execution frequency 50Hz, maximum number of iterations per cycle 5, number of parallel threads 16. Contact force model parameters: repulsive stiffness. =5000 N / m, sigmoid sharpness =20, attraction strength =5N (hand contact point) =0N (foot contact point), attractive force range σ=0.05m. Solver parameters: contact stiffness 2000 N / m, dissipation velocity 0.1 m / s, smoothing factor 0.005, ground friction coefficient. =0.7. Initial trust region radius. 0 = 0.1.

[0139] The underlying servo system uses PD control with a simulation step size of 0.002s (500Hz). PD gains for leg joints: hip roll Kp=250, Kd=5; hip yaw Kp=500, Kd=5; hip pitch Kp=250, Kd=5; knee Kp=250, Kd=5; ankle pitch Kp=20, Kd=2; ankle roll Kp=20, Kd=2. PD gains for arm joints: shoulder pitch Kp=10, Kd=1; shoulder roll Kp=50, Kd=1; shoulder yaw Kp=10, Kd=1; elbow Kp=10, Kd=1. Joint torque limits: Leg hip joint roll / pitch 180Nm, hip joint yaw 100Nm, knee joint 180Nm, ankle joint 36Nm; Arm shoulder joint pitch 100Nm, shoulder joint roll 50Nm, shoulder joint yaw 39Nm, elbow joint 50Nm.

[0140] In a preferred embodiment, a negative pressure adsorption device is installed on the handpiece. The robot is placed 0.48m above the wall, where the wall's coefficient of friction is... wall =0.5. The inverse kinematics of the handpiece are solved using the damped least squares method, with a maximum of 16 iterations, a step size of 0.7, and a damping coefficient of 5e-3. The system continuously collects data on the robot's center of mass position, velocity, attitude, and the contact distance between the handpiece and the wall. Based on the desired walking speed, the planning module generates the next foot placement position based on LIP dead zone feedback, and simultaneously generates the contact trajectory of the right arm extending towards the wall, ensuring that the handpiece maintains contact with the wall during walking.

[0141] The prediction module simplifies the robot into a single rigid body model, jointly optimizing the center of mass trajectory and contact pull within the prediction time window, and allowing the generation of negative normal force to form a stabilizing pull when the contact distance is close to zero. The whole-body mapping module, while prioritizing center of mass balance, transforms the prediction results into joint angle targets and suction cup negative pressure targets. The execution layer drives the robot to complete stable walking according to the targets, and the feedback results are used for the next cycle update.

[0142] Example 2: Stability control of handrail climbing in sloping environments

[0143] In the simulation verification scenario, the slope gradient =15°, coefficient of ground friction ground =0.4, coefficient of friction of the handrail bar =0.6. The gait planner adjusts the terrain height function based on the slope parameters: z terrain (x,y) = z origin + * tan( ),in This represents the horizontal distance along the slope. CIMPC parameter adjustments: contact stiffness increased to 5000 N / m to accommodate variations in the slope's normal force, smoothing factor reduced to 0.005 to improve accuracy, and friction coefficient set to 0.7. The desired robot walking speed is reduced to v. des =0.15 m / s. The attractive force parameter k at the hand contact point. a =0.5N, =0.005m. During the uphill climb, alternating hand grips on the poles create a pulling force assist. In contrast, the traditional CIMPC (hand grip) method uses... a =0, tension not allowed) under the same slope conditions, due to foot friction ground *mg*cos( ) is insufficient to completely offset the slope gravity component mg*sin( )(exist =0.4, At 15°, the safety margin is only approximately *cos( )-sin( (≈0.13), traditional methods result in significant slippage and center of gravity shift, while this invention compensates for insufficient friction by using hand-held suction force, thus maintaining stable climbing.

[0144] This invention also verifies the system's robustness to external disturbances. In simulations, a random disturbance generation module applies a continuous generalized force disturbance at a specified time (e.g., a horizontal thrust of 100N applied for 0.5s at t=2.0s) to verify the control system's recovery capability under disturbance. The gait planner rapidly adjusts the foot position through the LIP dead zone feedback mechanism, the CIMPC re-optimizes the centroid trajectory in the next prediction window, and the WBC automatically increases the balance task weight during the disturbance. Simulation results show that the system can recover to normal walking state within 2-3 steps after the disturbance ends.

[0145] The prediction module optimizes the center of mass momentum and hand contact force under dynamic equilibrium and friction constraints, enabling the robot to maintain a stable support area during slope walking. The full-body mapping module first ensures center of mass stability and anti-slip, then controls arm posture and suction pull. The execution layer adjusts motor output and suction cup negative pressure according to target instructions to complete continuous and stable climbing.

[0146] Example 3: Dynamic stability recovery under push-pull disturbance

[0147] External perturbations are applied to the robot in a simulated flat terrain scenario. These perturbations are implemented using the DisturbanceGenerator module, applying a generalized force perturbation lasting 0.5 seconds at t=2.0s. The force vectors are 10N in the x-direction, 5N in the y-direction, and 5N in the z-direction of the base. The perturbation force is directly injected into the multi-body plant model through the generalized force port. The system detects an increase in the lateral velocity of the center of mass using the EKF state estimator. The gait planner, based on LIP dead zone feedback, immediately corrects the next foot placement (computation time <0.1ms, purely analytical calculation). CIMPC re-optimizes the center of mass trajectory within the next 20ms control cycle and generates the arm contact trajectory for lateral wall support to provide restoring tension.

[0148] The prediction module calculates the new center-of-mass trajectory and lateral contact tension within the prediction window, enabling the system to anticipate disturbances and restore balance. The whole-body mapping module translates the recovery action into leg support adjustment and arm tension output. After the execution layer completes the action, it feeds back the stable state to the prediction module, achieving continuous disturbance-resistant stable control.

[0149] Example 4: Walking with support from both sides in a narrow passage

[0150] In a passageway 0.8m wide (slightly larger than the robot's shoulder width of 0.6m), the robot's two arms simultaneously make contact with the left and right walls. K is set at the contact points on both the left and right hands. a =5N Gaussian attraction parameter. CIMPC simultaneously optimizes the contact force sequence for four contact points (both feet + both hands), and solves the left and right arm targets separately using inverse kinematics of the hands, employing the damped least squares method (damping coefficient 5e-3). The system acquires the contact distance and adsorption pressure on both sides in real time, and coordinates the contact timing of both arms through the planning module.

[0151] The prediction module simultaneously optimizes the center-of-mass trajectory and left and right contact pull forces, enabling the robot to maintain a centered posture and walk stably in confined spaces. The full-body mapping module assigns different pull targets to the left and right arms and coordinates the gait of the legs. The execution layer drives the robot to smoothly pass through the passage according to the targets.

[0152] Example 5: Tensile-assisted stabilization during heavy-load handling

[0153] When the robot carries a 5kg load, its center of mass shifts forward by approximately 0.05m, causing its ZMP (Zero Motion Point) to approach the boundary of the supporting polygon. CIMPC incorporates the load inertia into the mass parameters of the single rigid body model (increasing the total mass to approximately 50kg), re-optimizes the center of mass trajectory within the prediction window, and automatically reduces the walking speed to v. des =0.1 m / s. After the system detects the shift in the center of gravity, the gait planner generates the contact target of the handrail in front, and the hand contact force parameter k. a =5N provides tensile strength to compensate for the instability caused by the load.

[0154] The prediction module optimizes the center of mass momentum and contact tension while taking into account the effects of load inertia to avoid the risk of tipping over. The whole-body mapping module prioritizes ensuring balance, then coordinates arm tension and gait, and the execution layer completes the coordinated control of handling and stability.

[0155] Example 6: Adaptive Stabilization Control under Continuous Switching of Complex Terrain

[0156] The robot passes through the following sections in sequence: (a) flat ground ( =0.7, no hand contact required, k a =0), (b) 15° slope section ( =0.4, activate single-arm pulling assistance, k a =5N), (c) Narrow passage with a width of 0.8m (with double-arm pulling assistance, left and right hands k a =5N). Environment switching is achieved through the terrain height function z. terrain (x,y) = z origin + *tan( Automatic detection: CIMPC switches contact strategies and constraints within one control cycle (20ms) after detecting terrain changes. The gait planner automatically adjusts the desired speed based on terrain parameters: 0.25m / s on flat ground → 0.15m / s on slopes → 0.1m / s on passageways.

[0157] The prediction module switches constraints and optimizes weights in real time for different terrains, enabling the robot to maintain stable control in various environments. The full-body mapping module dynamically adjusts task priorities, ensuring that balancing tasks are always given the highest priority, guaranteeing continuous and stable operation.

[0158] Example 7: Flexible Tensile Force Control Under Weak Contact Surfaces

[0159] Walking along a wall covered with a flexible material, the system detects the wall's flexibility by monitoring changes in contact distance and pressure. For flexible surfaces, the contact force model parameters are adjusted: the repulsive stiffness k is reduced. r Increase the attractive force range to 100 N / m (from 200 N / m). The distance was reduced to 0.01m (from 0.005m), making the contact force-distance curve smoother and avoiding sudden changes in contact force caused by flexible deformation. At the same time, the dissipation velocity in CIMPC was increased to 0.2m / s to enhance the damping effect.

[0160] The prediction module optimizes the center of mass trajectory and tension using an adjusted contact model, preventing the robot from becoming unstable due to flexible deformation. The whole-body mapping module reduces the rate of tension change and prioritizes center of mass stability, enabling the execution layer to achieve safe walking in a flexible environment.

[0161] Example 8: Adsorption Stability Control under High-Frequency Perturbation

[0162] In an environment with a 10Hz vibration source, the wall surface experiences minute vibrations with an amplitude of ±2mm. The system collects suction cup pressure fluctuations in real time at a sampling rate of 500Hz. After the EKF state estimator filters out high-frequency noise, CIMPC incorporates the vibration impact into the contact distance disturbance term within the prediction window. WBC increases the weight of the center-of-mass balancing task (the weight ratio is adjusted from the default 10:5:3:1 to 20:5:3:1), and the underlying PD controller maintains stable tension by rapidly adjusting the negative pressure output.

[0163] The whole-body mapping module increases the weight of the balance task and appropriately reduces the weight of the contact task. The execution layer maintains stable tension by quickly adjusting the negative pressure output, ultimately achieving stable walking in a vibration environment.

[0164] The technical effects obtained through simulation verification in the embodiments of the present invention include:

[0165] (1) Effectiveness of the tension contact model: The improved contact force function in Tension is generated near ≈0, with a peak value of k. a (Default 5N), effective range 3* ≈15mm. Foot contact point (k) a =0 degenerates into a traditional repulsion model, with the hand contact point set to k. a >0 enables attraction. The optimizer smoothly transitions between repulsion and attraction states, exhibiting good numerical convergence;

[0166] (2) Computational efficiency: Compared with the complete multi-rigid-body model (usually nq>100), the simplified single-rigid-body model (nq=27, nv=26) reduces the number of optimization variables by about 75%. With warm start and a 5-iteration limit, CIMPC runs at 50Hz (20ms cycle) and solves in parallel with 16 threads, meeting the real-time control requirements. The Drake kLagged discrete contact model is used in the simulation, with a simulation step size of 0.002s (500Hz).

[0167] (3) Gait planning response speed: The step size calculation for LIP dead zone feedback is an analytical form of matrix multiplication (2x2 matrix operation), with a single calculation time of <0.1ms. Compared with the HLIP regulator, which requires solving nonlinear optimizations for each step (typically taking 10-50ms), the response speed is improved by two orders of magnitude. The 7th-order Bezier trajectory interpolation of the swing foot is also an analytical calculation;

[0168] (4) Task Coordination: WBC allocates joint resources according to priority (center of mass balance > foot contact > hand contact > posture maintenance), with center of mass balance always being given priority. During the disturbance recovery process, WBC automatically reduces the weight of hand tasks to concentrate resources on maintaining balance, and normal walking resumes in 2-3 steps after the disturbance ends;

[0169] (5) Slope throughput capacity: On a 15° slope ( Under the condition of (k=0.4), traditional CIMPC (k a =0) because the safety margin for foot friction is only 0.13 ( *cos15°-sin15°) causes slippage. This invention compensates for the frictional gap of approximately 0.13*mg≈60N by using hand-held suction force, thus maintaining stable climbing.

[0170] (6) Joint torque constraints: The system has built-in joint torque limits (180 Nm for hip joint, 180 Nm for knee joint, and 36 Nm for ankle joint; 100 Nm for shoulder joint and 50 Nm for elbow joint). Both CIMPC and WBC meet the torque constraints during the optimization process.

[0171] It should be noted that embodiments of the present invention can be implemented in hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the above-described devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuitry such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of the above-described hardware circuitry and software, such as firmware.

[0172] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for tension control of a humanoid robot based on an improved contact model, characterized by, This method constructs a closed-loop control mechanism based on prediction and contact coordination, including the following steps: Step 1: Collect the robot's center of mass state and contact state. The center of mass state includes position, velocity, and orientation. The contact state includes the signed distance between each contact point. and adsorption pressure; Step 2: Based on the dynamics of the linear inverted pendulum (LIP) and the dead zone feedback mechanism, generate the landing target and the hand contact target according to the center of mass state and the desired motion requirements; Step 3: Within the prediction window, simplify the robot dynamics to a single rigid body model, and use the generalized coordinate trajectory q(0:N), generalized velocity trajectory v(0:N), and joint torque tau(0:N-1) as optimization variables. Under the conditions of satisfying dynamic constraints and friction constraints, jointly optimize the center of mass trajectory and contact force sequence. Step 4, in the prediction process, the contact force is calculated using an improved smooth continuity function: The first term is a sigmoid-based repulsive force term, and the second term is a Gaussian attractive force term; when the contact distance... When k ≈ 0, the attraction term generates a pulling force, simulating the suction cup adsorption effect; when the contact point is a normal foot support, let k a =0, the contact force function degenerates into a traditional model that only repels; Step 5: Based on the optimized centroid trajectory and contact force, map them into whole-body joint targets and adsorption pull targets according to task priority; Step 6: Send the joint target and the adsorption pull target to the execution layer to drive the robot's actions; feed back the actual state to the prediction layer to form a closed-loop update.

2. The method as described in claim 1, characterized in that, The calculation of the landing target is based on the feedforward-feedback structure of the linear inverted pendulum (LIP) model, and the step size calculation formula is: ,in Let T be the desired walking speed and T be the gait period. and These represent the position and velocity of the robot's center of mass at the moment of step change. and This is the corresponding reference value for the LIP model. and The dead zone feedback gain matrix is ​​used; the swing foot trajectory is generated by 7th-order Bezier curve interpolation.

3. The method as described in claim 1, characterized in that, The improved contact force function c( The value range of each parameter in () is as follows: repulsion stiffness k r Values ​​range from 500 to 5000 N / m, sigmoid sharpness. The value ranges from 5 to 20, representing the intensity of attraction, k. a Values ​​range from 0.1 to 10 N, attractive force range The value ranges from 0.001 to 0.01 m; In the CIMPC solver, the contact stiffness parameter is set to 2000–5000 N / m, the dissipation velocity to 0.05–0.2 m / s, the smoothing factor to 0.005–0.01, and the friction coefficient to... The value ranges from 0.5 to 0.7; the friction cone constraint is ||f t || <= * max(c( ), 0).

4. A humanoid robot pulling control device that implements the humanoid robot pulling control method based on an improved contact model as described in any one of claims 1-3, characterized in that, include: The state sensing unit is used to acquire the state of the center of mass and the contact state; The predictive optimization unit is used to jointly optimize the centroid trajectory and contact force based on a single rigid body dynamics model and a tensile contact model that allows negative normal forces. The task mapping unit is used to convert the center of mass trajectory and contact force into joint motion targets and adsorption pull targets according to task priority. An execution control unit is used to drive the robot's movements according to the joint motion target and the adsorption pull target; The feedback unit is used to feed back the execution status to the prediction and optimization unit to form a closed loop; In a preferred embodiment, a negative pressure adsorption device is installed at the end of a robot hand and includes a flexible sealed cavity, a negative pressure generating unit, and a pressure sensor to generate a controllable adsorption pull at the contact surface; the negative pressure generating unit may be an electrically driven vortex fan or a micro vacuum pump.

5. The apparatus as described in claim 4, characterized in that, The prediction and optimization unit simultaneously applies positional constraints and contact friction constraints between the centroid and the support region.

6. The apparatus as claimed in claim 4, characterized in that, The task mapping unit adopts a hierarchical priority constraint method, prioritizing the execution of balanced tasks, followed by landing tasks and contact tasks.

7. A whole-body control mapping method for humanoid robot pull control, implementing the humanoid robot pull control method based on an improved contact model as described in any one of claims 1-3, characterized in that, include: Receive the centroid trajectory and contact force generated by the prediction optimization unit; The centroid trajectory is converted into a balance constraint in the joint space; The contact force is converted into force targets at the hand and foot ends; According to the preset task priority order, the balance constraint is satisfied first, then the landing position constraint and contact posture constraint are satisfied, and finally the adsorption pull constraint is satisfied. Output the joint angle target, joint torque target, and adsorption pull target that satisfy the aforementioned priority constraint relationship.

8. The method as described in claim 7, characterized in that, The adsorption pull target is dynamically adjusted based on the deviation between the predicted contact force and the actual adsorption pressure.

9. The method as described in claim 7, characterized in that, The joint angle target and the joint torque target are solved simultaneously through a constraint optimization method.

10. The method as described in claim 7, characterized in that, The task priority is dynamically adjusted when the risk of robot instability increases, so that the priority of balancing tasks is higher than that of contact tasks and motion tasks.

Images