Robot and behavior planning method and control method thereof

Abstract

The embodiment of the invention discloses a robot and a behavior planning method and control method thereof, and the behavior planning method of the robot comprises the steps: determining a target grabbing pose of the robot for grabbing a target object based on an object pose of the target object; the target grabbing pose comprises a target base pose and a target gripper pose; based on the target base posture and the current base posture of the robot, a motion control strategy of mechanical legs of the robot is determined; the motion control strategy is used for controlling the mechanical legs to drive the robot to move, so that a base of the robot reaches a target base posture; on the basis of the target gripper pose and the current gripper pose of the robot, a gripping control strategy of a mechanical arm of the robot is determined; the grabbing control strategy is used for controlling the mechanical arm to drive the gripper of the robot to grab the target object at the target gripper pose under the condition that the base reaches the target base seat pose.

Description

Technical Field

[0001] The present disclosure relates to but is not limited to the field of robotics, and in particular to a robot and a behavior planning method and a control method thereof. Background Art

[0002] The movement and operation of legged robots integrate the leg movement and robotic arm operation of legged robots. Legged robots are widely used in home assistance, urban maintenance, disaster relief, and autonomous on-site operations. In order to improve the smoothness of leg movement and the accuracy of robotic arm operation in practical applications, in related technologies, the grasping detection results are applied to the robot's arm motion planning. This method ignores the coordination between the legged robot body and arm operation. Summary of the invention

[0003] In view of this, the embodiments of the present disclosure at least provide a robot and a behavior planning method and a control method thereof.

[0004] The technical solution of the embodiment of the present disclosure is implemented as follows:

[0005] The present disclosure provides a robot behavior planning method, including:

[0006] Based on the object posture of the target object, determining the target grasping posture of the robot grasping the target object; the target grasping posture includes the target base posture and the target gripper posture;

[0007] Based on the target base seat posture and the current base seat posture of the robot, a motion control strategy of the robot's mechanical legs is determined; the motion control strategy is used to control the mechanical legs to drive the robot to move so that the robot's base reaches the target base seat posture;

[0008] The grasping control strategy of the robot's mechanical arm is determined based on the target gripper posture and the robot's current gripper posture; the grasping control strategy is used to control the mechanical arm to drive the robot's gripper to grasp the target object at the target gripper posture when the base reaches the target base posture.

[0009] The present disclosure provides a robot control method, including:

[0010] Obtaining the target grasping posture of the robot grasping the target object; the target grasping posture includes the target base posture and the target gripper posture;

[0011] Obtain the motion control strategy of the robot's mechanical legs;

[0012] Controlling the mechanical legs to drive the robot to move so that the base of the robot reaches a target base posture;

[0013] Obtain the grasping control strategy of the robot's manipulator;

[0014] When the base reaches the target base posture, the robotic arm is controlled to drive the robot's gripper to grasp the target object in the target gripper posture.

[0015] The present disclosure provides a robot, including a planning module, a base, a gripper, a mechanical arm, and a mechanical leg, wherein:

[0016] The planning module is used to determine the target grasping posture of the robot to grasp the target object based on the object posture of the target object; the target grasping posture includes the target base seat posture and the target gripper posture; based on the target base seat posture and the current base seat posture of the robot, the motion control strategy of the robot's mechanical legs is determined; the motion control strategy is used to control the mechanical legs to drive the robot to move so that the base of the robot reaches the target base seat posture; based on the target gripper posture and the current gripper posture of the robot, the grasping control strategy of the robot's mechanical arm is determined; the grasping control strategy is used to control the mechanical arm to drive the robot's gripper to grasp the target object with the target gripper posture when the base reaches the target base seat posture.

[0017] The present disclosure provides a robot behavior planning device, including:

[0018] A first determination module is used to determine a target grasping posture of the robot grasping the target object based on the object posture of the target object; the target grasping posture includes a target base posture and a target gripper posture;

[0019] The second determination module is used to determine the motion control strategy of the robot's mechanical legs based on the target base seat posture and the current base seat posture of the robot; the motion control strategy is used to control the mechanical legs to drive the robot to move so that the base of the robot reaches the target base seat posture;

[0020] The third determination module is used to determine the grasping control strategy of the robot's mechanical arm based on the target grasper posture and the current grasper posture of the robot; the grasping control strategy is used to control the mechanical arm to drive the robot's grasper to grasp the target object with the target grasper posture when the base reaches the target base posture.

[0021] The present disclosure provides a control device for a robot, comprising:

[0022] The first acquisition module is used to acquire the target grasping posture of the robot grasping the target object; the target grasping posture includes the target base posture and the target gripper posture;

[0023] The second acquisition module is used to acquire the motion control strategy of the robot's mechanical legs;

[0024] The first control module is used to control the mechanical legs to drive the robot to move so that the base of the robot reaches a target base posture;

[0025] A third acquisition module is used to acquire a grasping control strategy of the robot's mechanical arm;

[0026] The second control module is used to control the mechanical arm to drive the robot's gripper to grasp the target object in the target gripper posture when the base reaches the target base posture.

[0027] An embodiment of the present disclosure provides a computer device, including a memory and a processor, wherein the memory stores a computer program that can be run on the processor, and when the processor executes the program, some or all of the steps in the above method are implemented.

[0028] An embodiment of the present disclosure provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, some or all of the steps in the above method are implemented.

[0029] An embodiment of the present disclosure provides a computer program product, including a computer program or instructions. When the computer program or instructions are executed by a processor, some or all of the steps in the above method are implemented.

[0030] In the disclosed embodiment, firstly, according to the object posture of the target object, the target grasping posture of the robot grasping the target object is determined; the target grasping posture includes the target base posture and the target gripper posture; secondly, according to the target base posture and the current base posture of the robot, the motion control strategy of the robot's mechanical legs is determined; the motion control strategy can be used to control the mechanical legs to drive the robot to move so that the base of the robot reaches the target base posture; finally, according to the target gripper posture and the current gripper posture of the robot, the grasping control strategy of the robot's mechanical arm is determined; the grasping control strategy can be used to control the mechanical arm to drive the robot's gripper to grasp the target object in the target gripper posture when the base reaches the target base posture. In this way, the motion control strategy of the mechanical legs is planned so that the robot approaches the target grasping posture, and when the base reaches the target base posture, the grasping control strategy of the mechanical arm is planned so that the gripper reaches the target gripper posture so that the robot reaches the target gripper posture, so that the robot moves smoothly and accurately grasps the target object.

[0031] It should be understood that the above general description and the following detailed description are merely exemplary and explanatory, and are not intended to limit the technical solutions of the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] The drawings herein are incorporated into the specification and constitute a part of the specification. These drawings illustrate embodiments consistent with the present disclosure and are used to illustrate the technical solutions of the present disclosure together with the specification.

[0033] Figure 1A schematic diagram of an implementation flow of a robot behavior planning method provided in an embodiment of the present disclosure;

[0034] Figure 2 A schematic diagram of an implementation flow of a robot control method provided in an embodiment of the present disclosure;

[0035] Figure 3 A schematic diagram of the structure of a robot provided in an embodiment of the present disclosure Figure 1 ;

[0036] Figure 4 A schematic diagram of the structure of a robot provided in an embodiment of the present disclosure Figure 2 ;

[0037] Figure 5 A schematic diagram of the structure of a robot provided in an embodiment of the present disclosure Figure 3 ;

[0038] Figure 6 A schematic diagram of the structure of a robot provided in an embodiment of the present disclosure Figure 4 ;

[0039] Figure 7 A schematic diagram of the structure of a robot provided in an embodiment of the present disclosure Figure 5 ;

[0040] Figure 8 A schematic diagram of a scene in which a robot grabs different objects in the real world provided by an embodiment of the present disclosure;

[0041] Fig. 9 A schematic diagram of a full body grabbing frame system provided by an embodiment of the present disclosure;

[0042] Fig.10 A schematic diagram of the implementation process of the whole body grabbing framework provided in the embodiment of the present disclosure;

[0043] Fig.11 A schematic diagram of the implementation process of the motion control module provided in the embodiment of the present disclosure;

[0044] Fig.12 A schematic diagram of the implementation process of the perception module provided in the embodiment of the present disclosure;

[0045] Fig.13 A schematic diagram of the implementation process of the operation module provided in the embodiment of the present disclosure;

[0046] Fig.14 A schematic diagram of the implementation process of the planning module provided in the embodiment of the present disclosure;

[0047] Fig.15 A schematic diagram of a robot grabbing a target object provided by an embodiment of the present disclosure;

[0048] Fig.16 A schematic diagram of an arm workspace and a whole body workspace provided in an embodiment of the present disclosure;

[0049] Fig.17 A schematic diagram of a data set used in a real-world test provided by an embodiment of the present disclosure;

[0050] Fig.18 A schematic diagram of the structure of a robot behavior planning device provided by an embodiment of the present disclosure;

[0051] Fig.19 A schematic diagram of the structure of a robot control device provided by an embodiment of the present disclosure;

[0052] Fig. 20 A hardware entity schematic diagram of a computer device provided in an embodiment of the present disclosure. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical solutions and advantages of the present disclosure clearer, the technical solutions of the present disclosure are further elaborated in detail below in conjunction with the drawings and embodiments. The described embodiments should not be regarded as limiting the present disclosure. All other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present disclosure.

[0054] In the following description, reference is made to “some embodiments”, which describe a subset of all possible embodiments, but it will be understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0055] The terms "first / second / third" involved are merely used to distinguish similar objects and do not represent a specific ordering of the objects. It is understandable that "first / second / third" can be interchanged with a specific order or sequence where permitted so that the embodiments of the present disclosure described herein can be implemented in an order other than that illustrated or described herein.

[0056] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which the present disclosure belongs. The terms used herein are only for the purpose of describing the disclosure and are not intended to limit the disclosure.

[0057] The present disclosure provides a robot behavior planning method. The present disclosure provides a robot behavior planning method. Figure 1 As shown, the method includes the following steps S101 to S103:

[0058] Step S101: based on the object posture of the target object, determining the target grasping posture of the robot grasping the target object; the target grasping posture includes a target base posture and a target gripper posture.

[0059] Here, the object pose includes the position and attitude of the target object.

[0060] In some embodiments, the position of the target object includes the position information of the target object in space. The posture of the target object includes the direction or angle information of the target object relative to the coordinate system. The robot includes a legged robot, for example, a quadruped robot. In the disclosed embodiment, the robot includes a mechanical leg and a mechanical arm, the mechanical leg is used to drive the robot to move smoothly and drive the base of the robot to move, and the mechanical arm is used to drive the gripper of the robot to move. The target base posture includes the position and posture of the base of the robot, and the movement of the mechanical leg can be controlled to make the base reach the target base posture; the target gripper posture includes the position and posture of the gripper of the robot, and the movement of the mechanical arm can be controlled to make the gripper reach the target gripper posture.

[0061] In some embodiments, a generalized directional reachability graph is used to support six degrees of freedom for base positioning, wherein the six degrees of freedom include three translational degrees of freedom (forward / backward, left / right displacement, up / down movement) and three rotational degrees of freedom (pitch, yaw, roll). The generalized directional reachability graph may represent candidate base seat poses of the six degrees of freedom relative to the tool center point frame.

[0062] Step S102: Based on the target base seat posture and the current base seat posture of the robot, determine the motion control strategy of the robot's mechanical legs; the motion control strategy is used to control the mechanical legs to drive the robot to move so that the base of the robot reaches the target base seat posture.

[0063] Here, the current base posture includes the current base position and the current base posture, and the target base posture includes the target base position and the target base posture.

[0064] In the disclosed embodiment, the posture of the base can be changed by controlling the movement of the robotic legs, so the movement of the robotic legs can be planned to move the base from the current base posture to the target base posture.

[0065] In some embodiments, a motion control strategy of the robotic legs is determined during the process of the base moving from the current base posture to the target base posture. The motion control strategy can be used to control the motion of the robotic legs so that the base moves to the target base position with the current base posture, and then control the motion of the robotic legs so that the base moves from the current base posture to the target base posture; or the motion control strategy can be used to control the motion of the robotic legs so that the base reaches the target base position with the target base posture.

[0066] Step S103: Determine the grasping control strategy of the robot's robotic arm based on the target gripper posture and the robot's current gripper posture; the grasping control strategy is used to control the robotic arm to drive the robot's gripper to grasp the target object at the target gripper posture when the base reaches the target base posture.

[0067] Here, the current gripper posture includes the current gripper position and the current gripper posture, and the target gripper posture includes the target gripper position and the target gripper posture.

[0068] In the disclosed embodiment, the gripper's posture can be changed by controlling the movement of the robotic arm, so the operation of the robotic arm can be planned so that the gripper moves from the current gripper posture to the target gripper posture.

[0069] In some embodiments, a grasping control strategy of the manipulator during the movement of the gripper from the current gripper posture to the target gripper posture is determined. The grasping control strategy can be used to control the movement of the manipulator so that the gripper moves to the target gripper position with the current gripper posture, and then control the movement of the manipulator so that the gripper moves from the current gripper posture to the target gripper posture; or, the grasping control strategy can be used to control the movement of the manipulator so that the gripper reaches the target gripper position with the target gripper posture.

[0070] In the disclosed embodiment, firstly, according to the object posture of the target object, the target grasping posture of the robot grasping the target object is determined; the target grasping posture includes the target base seat posture and the target gripper posture; secondly, according to the target base seat posture and the current base seat posture of the robot, the motion control strategy of the robot's mechanical legs is determined; the motion control strategy can be used to control the mechanical legs to drive the robot to move so that the base of the robot reaches the target base seat posture; finally, according to the target gripper posture and the current gripper posture of the robot, the grasping control strategy of the robot's mechanical arm is determined; the grasping control strategy can be used to control the mechanical arm to drive the robot's gripper to grasp the target object in the target gripper posture when the base reaches the target base seat posture. In this way, the motion control of the mechanical legs and the operation of the mechanical arms are combined by using the target base seat posture and the target gripper posture, so that the robot moves smoothly and accurately grasps the target object.

[0071] In some embodiments, before executing step S101 to determine the target grasping posture of the robot grasping the target object based on the object posture of the target object, the method may include the following steps S111 and S112:

[0072] Step S111: Acquire image information of the target object.

[0073] The image information may be image information collected by an image collection component disposed on the robotic arm.

[0074] Image information may include color image information and depth image information. Color image information provides visual features such as color, texture, shape, etc. of the target object; depth image information provides the spatial distance relationship between the target object and the robot.

[0075] Step S112: Based on the image information, the target object is identified and located to determine the object posture of the target object.

[0076] In some implementations, the image information is segmented to obtain the features of the target object, the depth of the target object is predicted based on the image information to obtain the spatial information of the target object, and the object position of the target object is determined based on the features of the target object and the spatial information of the target object.

[0077] In the disclosed embodiment, image information of the target object is obtained; the target object is identified and located based on the image information, and the object posture of the target object is determined. In this way, an accurate object posture can be obtained for subsequent prediction of the target grasping posture.

[0078] In some embodiments, the image information includes an infrared image and a color image; the above step S112 identifies and locates the target object based on the image information and determines the object posture of the target object, which may include the following steps S121 to S123:

[0079] Step S121: Segment the target object in the color image to obtain a segmentation result.

[0080] In some embodiments, a target tracking model (Track Anything Model, TAM) is used to process the color image. TAM may include a segmentation model (Segment Anything Model, SAM) and a memory module (XMem). The color image is segmented using SAM to obtain an initial mask, and XMem is used for tracking to obtain a continuous real-time mask during the robot's motion to obtain a segmentation result.

[0081] Step S122: Determine the depth of the target object based on the infrared image and the color image.

[0082] In some embodiments, the depth of the target object is predicted using a learning-based infrared spot approach using a generalized grasping technique (ASGrasp) based on the infrared image and the color image.

[0083] Step S123: Determine the object pose of the target object based on the segmentation result and the depth of the target object.

[0084] In some implementations, the object pose of the target object is determined by combining the features of the target object in the segmentation result and the spatial information contained in the depth.

[0085] In the disclosed embodiment, the target object in the color image is segmented to obtain a segmentation result; the depth of the target object is determined based on the infrared image and the color image; and the object posture of the target object is determined based on the segmentation result and the depth of the target object. In this way, the accurate object posture is obtained using the segmentation result and the depth of the target object.

[0086] In some embodiments, the above step S101 determines the target grasping posture of the robot grasping the target object based on the object posture of the target object, and may include the following steps S131 and S132:

[0087] Step S131: Based on the object posture, determine the target gripper posture for the gripper to grasp the target object, that is, determine the target gripper posture for the gripper to grasp the target object according to the position and posture of the target object.

[0088] Step S132: Determine the target base seat pose based on the target gripper pose, that is, plan the target base seat pose according to the target gripper pose.

[0089] In the disclosed embodiment, the target base seat posture is a base seat posture that satisfies the target gripper posture, so that the robot arm and the body move in coordination.

[0090] In some embodiments, the above step S132 determines the target base posture based on the target gripper posture, which may include the following steps S141 and S142:

[0091] Step S141: Based on the target gripper posture, determine a plurality of candidate base seat postures for enabling the gripper to reach the target gripper posture.

[0092] Based on the target gripper posture, the generalized oriented reachability graph (GORM) is used to obtain multiple candidate base seat postures, and the base seat postures that exceed the motion range, collide with the environment, and are below the reachability threshold are removed from the multiple candidate base seat postures; the distribution of multiple candidate bases in the world coordinate system is calculated by homogeneous inverse transformation (RM). In this way, the base seat posture in the workspace can be optimized to obtain high-quality candidate base seat postures.

[0093] Step S142: Based on the current base seat pose of the robot, determine a target base seat pose having the shortest distance from the current base seat pose from a plurality of candidate base seat poses.

[0094] In some implementations, a distance distribution is calculated based on the distances between the current base seat pose and each candidate base position, and a minimum distance may be selected from the distance distribution as the target base seat pose.

[0095] In the disclosed embodiment, the robot may be encouraged to approach a target base seat posture that is the smallest distance from the current base seat posture, thereby improving the efficiency of the robot's movement.

[0096] In some embodiments, the above step S102 determines the motion control strategy of the robot's mechanical legs based on the target base seat posture and the current base seat posture of the robot, and may include the following steps S151 and S152:

[0097] Step S151: Based on the target base posture and the current base posture, determining the movement speed of the mechanical leg and the angles of each leg joint when the base moves from the current base posture to the target base posture.

[0098] The above-mentioned planning of the movement of the mechanical leg may include planning the movement speed of the mechanical leg and the angles of each leg joint, and may also include parameters such as angular velocity and gait variables.

[0099] In some embodiments, the movement speed of the robotic legs and the angles of each leg joint are planned according to the target base posture and the current base posture when the base moves from the current base posture to the target base posture.

[0100] Step S152: Determine the motion control strategy based on the movement speed of the mechanical leg and the angles of each leg joint.

[0101] In the disclosed embodiment, the moving speed of the mechanical legs and the angles of each leg joint are determined according to the target base posture and the current base posture during the movement of the base from the current base posture to the target base posture; the motion control strategy is determined according to the moving speed of the mechanical legs and the angles of each leg joint. In this way, the robot can be moved smoothly according to the motion control strategy.

[0102] In some embodiments, the target base posture includes a target base position and a target base posture; the above step S151 determines the movement speed of the mechanical leg and the angles of each leg joint during the movement of the base from the current base posture to the target base posture based on the target base posture and the current base posture, and may include the following steps S161 and S162:

[0103] Step S161: Based on the current base position and the target base position of the robot, determine the movement speed of the robot legs and the angles of each leg joint when the base moves from the current base position to the target base position.

[0104] First, the movement speed of the robotic legs and the angles of each leg joint are planned when the base moves from the current base position to the target base position; then, the angles of each leg joint of the robotic legs are planned when the base moves from the first posture of the base at the target base position to the target base posture.

[0105] Step S162: Based on the first posture of the base when the robot is at the target base position and the target base posture, determine the leg joint angles of the robot legs during the process of the base moving from the first posture to the target base posture.

[0106] In the disclosed embodiment, the base can be quickly moved to a target base position, and the posture of the base can be adjusted at the target base position using the shortest path.

[0107] In some embodiments, the target base posture includes a target base position and a target base posture; the above step S151 determines the movement speed of the mechanical leg and the angles of each leg joint during the movement of the base from the current base posture to the target base posture based on the target base posture and the current base posture, and may include the following step S163:

[0108] Step S163: Based on the current base posture and the target base posture, determine the movement speed of the mechanical leg and the leg joint angles in the process of the base moving from the current base posture to the target base position with the target base posture.

[0109] In some implementations, the posture of the base is adjusted during the process of the base moving from the current base position to the target base position, so that the base reaches the target base posture with the target base posture.

[0110] In the disclosed embodiment, the posture of the base can be adjusted to the target base posture when the base is approaching the target object, thereby improving efficiency.

[0111] In some embodiments, the step S102 described above of controlling the mechanical legs to drive the robot to move so that the base of the robot reaches the target base posture may include the following steps S171 and S172:

[0112] Step S171: Based on the motion control strategy, a set of leg motion instructions for the robotic leg is generated.

[0113] Among them, each leg motion instruction in the leg motion instruction set can be a five-dimensional instruction, including x-axis, y-axis, z-axis, pitch, and yaw. In order to make the robot move more smoothly, the five-dimensional instruction does not include rolling.

[0114] In some embodiments, for the stability of the robot and to prevent erroneous behavior, the sampling of the range of the leg movement command set is restricted according to the height and pitch angle: θ~f,v x ,v y ,ω~g(h,θ), where the pitch angle θ satisfies the preset constraint f, and the moving speed of the x-axis is v x , the moving speed of the y-axis is v y , the angular velocity ω satisfies the preset constraint g(h,θ).

[0115] Step S172: Based on the leg movement instruction set, control the mechanical legs to drive the robot to move so that the base moves from the current base sitting posture to the target base sitting posture.

[0116] In order to make the robot walk naturally, the teacher-student model is used to control the mechanical legs to drive the robot to move based on the leg motion instruction set so that the base moves from the current base posture to the target base posture; the teacher observation in the teacher-student model includes proprioception and privileged information; proprioception includes at least one of the following: the robot's current leg motion instruction, gravity vector, angular velocity, joint angle, previous leg motion instruction and gait variable; privileged information includes at least one of the following: the speed of the robot's base relative to the ground, the friction coefficient of each contact surface, the mass parameter and the motor strength.

[0117] In some embodiments, relevant parameters of the robot's leg motion are sampled in spherical coordinates with constant height and pitch angles, and randomization techniques and observation noise are used to determine relevant parameters of the robot's motion to reduce the gap between the simulated motion and the actual motion of the robot.

[0118] In the disclosed embodiment, a reward function is defined to track the set of leg movement instructions; the speed, pitch angle, and angular velocity in the vertical direction (up and down directions) of the body are penalized to achieve efficient and natural walking.

[0119] In some embodiments, after a robust teacher model is trained, it is refined into a student model, where the teacher model generates a latent vector using a privileged information encoder Then generate actions through a multi-layer perceptron (MLP) The student replaces the encoder with a gated recurrent unit (GRU), which receives proprioceptive observation data o t prop And generate the latent vector lt , l t With o t prop After connection, MLP is used to generate action a t .

[0120] Yes t and a t For supervision, the loss function is defined as follows:

[0121]

[0122] In the disclosed embodiment, according to the motion control strategy, a leg motion instruction set of the mechanical leg is generated; based on the leg motion instruction set, the mechanical leg is controlled to drive the robot to move so that the base moves from the current base posture to the target base posture. In this way, accurate control of the movement of the mechanical leg is achieved according to the leg motion instruction set.

[0123] In some embodiments, the above step S103 determines the grasping control strategy of the robot's mechanical arm based on the target gripper posture and the current gripper posture of the robot, and may include the following steps S181 to S183:

[0124] Step S181: determining the motion trajectory of the robot arm in the process of the gripper moving from the current gripper posture to the target gripper posture.

[0125] The motion planning method is used to plan the motion trajectory of the robot arm when the gripper moves from the current gripper posture to the target gripper posture, and the full-body reinforcement learning (RL) strategy is used to solve the control error problem in the end-effector control.

[0126] The motion of the image acquisition component set on the robot arm can be constrained to a tracker ball to minimize motion blur and prevent tracking loss. In addition, the tracker ball is defined using a reachability map (RM) to ensure that the image acquisition component only moves within the dexterous workspace. Within the dexterous workspace, the motion trajectory is generated using a motion planner.

[0127] Step S182: Determine the angles of each arm joint of the robotic arm based on the motion trajectory. That is, the angles of each arm joint of the robotic arm can be determined based on the motion trajectory of the robotic arm.

[0128] Step S183: Determine the grasping control strategy based on the arm joint angles of the robotic arm.

[0129] In the disclosed embodiment, the angles of each arm joint of the robotic arm are determined according to the movement trajectory of the robotic arm during the movement of the gripper from the current gripper position to the target gripper position; and the grasping control strategy is determined based on the angles of each arm joint of the robotic arm. In this way, the gripper can be moved to the target gripper position according to the grasping control strategy to accurately grasp the target object.

[0130] In some embodiments, the above step S103 of controlling the mechanical arm to drive the gripper of the robot to grasp the target object in the target gripper posture may include the following steps S191 and S192:

[0131] Step S191: Based on the grasping control strategy, generate an arm grasping instruction set for the robotic arm.

[0132] Each arm grasping instruction in the arm grasping instruction set can be a six-dimensional instruction, including x-axis, y-axis, z-axis, pitch, yaw, and roll.

[0133] Step S192: Based on the arm grasping instruction set, control the robotic arm to drive the gripper to move from the current gripper posture to the target gripper posture to grasp the target object.

[0134] In the disclosed embodiment, according to the grasping control strategy, an arm grasping instruction set of the robot arm is generated; based on the arm grasping instruction set, the robot arm is controlled to drive the gripper to move from the current gripper position to the target gripper position to grasp the target object. In this way, accurate control of the operation of the robot arm is achieved according to the arm grasping instruction set.

[0135] In some embodiments, when the base reaches the target base posture as described in step S103 above, controlling the mechanical arm to drive the gripper of the robot to grasp the target object in the target gripper posture may include the following step S193:

[0136] Step S193: When the distance between the current base seat posture and the target base seat posture is less than or equal to a preset distance threshold, control the gripper to grasp the target object with the target gripper posture.

[0137] During the movement of the robot, the movement of the robot is tracked, and the reachability of the current grasping posture is calculated using RM. When the reachability is less than or equal to the preset distance threshold, the state of the gripper is switched to the grasping stage and the target object is grasped with the target gripper posture.

[0138] In the disclosed embodiment, when the distance between the current base seat posture and the target base seat posture is less than or equal to the preset distance threshold, the gripper is controlled to grasp the target object in the target gripper posture. In this way, when the gripper can grasp the target object, it automatically switches to the grasping stage and controls the gripper to grasp the target object, thereby improving the efficiency of the gripper in grasping the target object.

[0139] The present disclosure provides a robot control method, and the present disclosure provides a robot control method implementation flow chart, such as Figure 2 As shown, the method includes the following steps S201 to S205:

[0140] Step S201: obtaining a target grasping posture of the robot grasping the target object; the target grasping posture includes a target base posture and a target gripper posture.

[0141] Step S202: Acquire a motion control strategy of the robot's mechanical legs based on the target base posture.

[0142] Step S203: Based on the motion control strategy, control the mechanical legs to drive the robot to move, so that the base of the robot reaches the target base posture.

[0143] Step S204: obtaining a grasping control strategy of the robot's mechanical arm based on the target gripper posture.

[0144] Step S205: Based on the grasping control strategy, when the base reaches the target base posture, the mechanical arm is controlled to drive the gripper of the robot to grasp the target object in the target gripper posture.

[0145] The above robot control method can be understood by referring to the explanation of the above robot behavior planning method embodiment, and has similar beneficial effects as the above robot behavior planning method embodiment.

[0146] The present disclosure provides a robot, such as Figure 3 As shown, the robot 300 includes a planning module 301, a base 311, a gripper 312, a mechanical arm 313 and a mechanical leg 314, wherein:

[0147] The planning module 301 is used to determine the target grasping posture of the robot 300 for grasping the target object based on the object posture of the target object; the target grasping posture includes a target base seat posture and a target gripper posture; based on the target base seat posture and the current base seat posture of the robot 300, the motion control strategy of the mechanical leg 314 of the robot 300 is determined; the motion control strategy is used to control the mechanical leg 314 to drive the robot 300 to move so that the base 311 of the robot 300 reaches the target base seat posture; based on the target gripper posture and the current gripper posture of the robot, the grasping control strategy of the mechanical arm 313 of the robot 300 is determined; the grasping control strategy is used to control the mechanical arm 313 to drive the gripper 312 of the robot 300 to grasp the target object with the target gripper posture when the base 311 reaches the target base seat posture.

[0148] In some embodiments, Figure 4 As shown, the robot 300 further includes a perception module 302, wherein:

[0149] The perception module 302 is used to obtain image information of the target object; based on the image information, identify and locate the target object, and determine the object posture of the target object.

[0150] In some embodiments, Figure 5 As shown, the robot 300 also includes a motion control module 303, wherein:

[0151] The motion control module 303 is used to generate a set of leg motion instructions for the mechanical leg 314 based on the motion control strategy; based on the set of leg motion instructions, control the mechanical leg 314 to drive the robot 300 to move so that the base 311 moves from the current base posture to the target base posture.

[0152] In some embodiments, Figure 6 As shown, the robot 300 further includes an operation module 304, wherein:

[0153] The operation module 304 is used to generate an arm grasping instruction set for the robotic arm 313 based on the grasping control strategy; based on the arm grasping instruction set, control the robotic arm 313 to drive the gripper 312 to move from the current gripper posture to the target gripper posture to grasp the target object.

[0154] In some embodiments, Figure 7 As shown, the robot 300 further includes an image acquisition component 321, wherein:

[0155] The image acquisition component 321 is arranged on the mechanical arm 313;

[0156] The image acquisition component 321 is used to obtain image information of the target object.

[0157] The following describes the application of the embodiments of the present disclosure in actual scenarios.

[0158] Legged robots with advanced manipulation capabilities have significant potential for improving household chores and urban maintenance. Despite progress in developing robust motion control and precise manipulation methods, it remains challenging to seamlessly integrate these methods into a unified full-body control system for practical applications.

[0159] The locomotion and manipulation of quadrupedal robots, which integrates legged locomotion and arm manipulation, is an important area of ​​research due to its wide range of potential applications, including in home assistance, urban maintenance, disaster relief, and autonomous field operations. Recent advances in reinforcement learning (RL) have enabled the development of complete policies for full-body locomotion and manipulation, enabling robots to perform tasks that seamlessly coordinate movement and interaction with objects. Although end-to-end reinforcement learning has significantly improved locomotion capabilities, mobile manipulation remains challenging due to the increased dimensionality of actions and the complex physical interactions involved. These challenges often result in mobile manipulation policies with limited accuracy and generalization, especially when grasping objects of different shapes, sizes, and materials. Scenarios of robots grasping different objects in the real world can be illustrated by Figure 8 As shown, the robot grasps an object in a cluttered environment 801 , the robot grasps an object at a first height 802 , and the robot grasps an object at a second height 803 .

[0160] To improve the performance and versatility of full-body grasping systems, we draw inspiration from the success of various grasp detection techniques. These methods have shown strong performance in detecting grasp poses of a variety of unseen objects (including challenging materials such as transparent or mirrored surfaces) in cluttered environments. By combining grasp pose detection with motion planning, these methods can achieve very high accuracy, with grasp success rates of over 90% for arbitrary objects in tabletop environments.

[0161] This inspired the idea of ​​combining legged motion with grasp detection to achieve high-performance, high-versatility motion and manipulation, thus achieving the best of both worlds. However, this combination is not easy, and it is not enough to directly apply grasp detection results to the arm motion planning of legged robots, ignoring the coordination required between the legged robot's body and arm motion.

[0162] In the related art, legged robot control methods rely on control-based methods to complete basic motion tasks, such as model predictive control (MPC). These methods perform well in constrained and controllable environments, but usually require precise modeling and manual debugging. With the rise of deep neural networks, learning-based methods have become more robust and adaptable, allowing robots to complete a wide range of movements with less human intervention. Such learning methods have made significant progress in robust motion control, agile motor skills, dynamic jumping, fall recovery, complex parkour movements, and performance in complex terrain and confined spaces.

[0163] In terms of legged robot motion manipulation, modularization and unified approaches are the two main development directions. The modular approach divides the motion and operation system into independent components, usually managing the motion part through an off-the-shelf controller, and optimizing the operation technology according to the specific task. This approach simplifies the design process and enhances the flexibility and stability of the system. In contrast, the unified approach is committed to achieving seamless motion manipulation through full-body control, such as controlling the leg joints and the robotic arm at the same time through a unified strategy to improve the overall performance.

[0164] Control-based methods are effective in specific environments, but they require precise modeling and complex manual debugging, which limits the adaptability of robots in changing environments. Although learning-based methods have made progress in many aspects, the problem they face is how to organically combine motion and manipulation capabilities.

[0165] Although the modular approach enhances the flexibility of the system, it treats the arm and body as independent components, limiting the expansion of the workspace and the effect of whole-body control. Although the unified approach excels in seamless control, it often relies on remote operation, resulting in insufficient autonomy. At the same time, early unified methods also have problems such as limited arm posture and low tracking accuracy, which have not been completely solved even with the introduction of a two-stage strategy. In addition, the SLAM-based whole-body control method requires a large amount of task data and external processing, which limits the motion manipulation capabilities. Although existing studies have integrated high-level task planning with low-level control, their grasping accuracy and task generalization capabilities are still insufficient, and require a lot of training time.

[0166] Based on the above description, a generalizable quadrupedal whole-body grasping framework (QuadWBG) for legged robots is proposed in the embodiments of the present disclosure using a modular and unified approach. The modular whole-body grasping framework is used for a robust and generalizable whole-body motion manipulation controller based on a single arm-mounted camera. Through reinforcement learning (RL), a robust low-level strategy for executing commands in a five-dimensional (5D) space is implemented, and a high-level strategy for perceptual grasping based on a "generalized directional reachability map" (GORM) is introduced. The whole-body grasping framework achieved an advanced one-time grasping accuracy of 89% in a real environment, and successfully completed challenging tasks including grasping transparent objects. Through extensive simulation and real-world experiments, it was demonstrated that the system can effectively manage a wide range of workspaces from the ground to above the body, and perform a variety of whole-body motion manipulation tasks.

[0167] The whole body grasping framework is a modular system, such as Fig. 9 As shown, it is divided into four key components: motion control module 303, perception module 302, operation module 304 and planning module 301. First, the five-dimensional instructions are tracked through reinforcement learning strategies, so as to achieve the robust mobility of the robot through motion control module 303. At the same time, the perception module 302 generates real-time object masks and grasping postures to guide the operation module to switch from tracking to grasping in the body frame. The planning module 301 is trained based on GORM to optimize the position of the robot base and improve the grasping performance. All strategies are trained in a simulation environment using the proximal policy optimization algorithm (PPO).

[0168] The full-body grasping framework is built on a quadruped robot (such as the Unitree B1), equipped with a robotic arm (such as the Unitree Z1 arm), a motorized gripper (such as the Robotiq gripper), and a camera mounted on the wrist (such as the Intel RealSense D415 camera).

[0169] The flowchart of each module in the modular system is as follows: Fig.10 As shown, it may include steps S1001 to S1004:

[0170] Step S1001: the perception module 302 receives the image signal, obtains the real-time grasping posture based on the neural network, and sends the real-time grasping posture to the planning module 301 and the operation module 304;

[0171] Step S1002: the planning module 301 receives the real-time grasping posture, determines the motion control instructions of the robot leg based on the neural network, and sends the motion control instructions to the motion control module 303;

[0172] Step S1003: the operation module 304 receives the real-time grasping posture and controls the arm operation based on the neural network;

[0173] Step S1004: the motion control module 303 receives the motion control instruction and controls the robot motion based on the neural network.

[0174] The following describes the four modules respectively:

[0175] (1) Motion control module;

[0176] The specific implementation process of the motion control module can be as follows Fig.11 As shown, it may include steps S1101 to S1105:

[0177] Step S1101: Generate a teacher motion control strategy;

[0178] Generate teacher movement control strategies based on privileged information and proprioception.

[0179] Step S1102: generating a student movement control strategy;

[0180] Based on proprioception, the teacher's motion control strategy is used for online supervised learning to generate the student's motion control strategy.

[0181] Step S1103: Experiment in a simulation environment.

[0182] Experiment in a simulation environment based on the target motor position determined by the teacher motion control strategy and the student motion control strategy.

[0183] Flexible motion strategies are essential for achieving high precision in whole-body motion control systems. To this end, the disclosed embodiment adopts a teacher-student architecture and expands the five-dimensional instruction set to include the pitch angle and height of the robot body center. To ensure stability and avoid undesirable behavior, the instruction range is constrained according to the pitch angle and height: θ~f,v x ,v y ,ω~g(h,θ), it can be seen that the pitch angle θ satisfies the f constraint, and the moving speed of the x-axis is v x , the moving speed of the y-axis is v y , the angular velocity ω satisfies the g(h,θ) constraint.

[0184] Observation of teacher behavior consists of proprioceptive and privileged information proprioception Contains instructions (corresponding to the leg movement instructions in the above embodiment) Gravity vector Angular velocity Joint position and velocity (i.e. joint angle) qt ,q t ∈R 18 , previous action a t-1 ∈R 12 and gait variables Privileged information includes parameters that cannot be accessed in real time, such as the speed of the robot's base relative to the ground, the friction coefficient of each contact surface, mass parameters, and motor strength.

[0185] In the disclosed embodiment, a reward function is defined to track instructions to achieve efficient and natural walking; the vertical velocity of the body, the angular velocity of the roll and pitch angles are penalized; the Raiber theuristic is used to determine the position of the robot's feet and the gap between the feet to prevent stumbling, as well as the smoothness of the action to achieve natural movement. The robot's gait is adjusted and the energy reward is given to obtain a more stable and smooth gait.

[0186] The motion strategy needs to accurately track 5D instructions under the disturbances generated when the operating module performs arm tracking and grasping the target object. To this end, the embodiment of the disclosure adopts a sampling method based on spherical coordinates with invariant height and pitch angles. In addition, wide domain randomization technology and observation noise are applied to reduce the gap between simulation and reality in robot dynamics (see Table 1). Among these strategies, increasing proprioceptive delay has been shown to be key to preventing unstable motion caused by onboard computing delays.

[0187] Table 1 shows the randomly set parameters. In Table 1, the minimum value of ground friction is 0.1 and the maximum value is 3.0; the minimum value of basic load mass is 0.0 and the maximum value is 20.0, in kilograms (kg); the body center of mass (x) is the position of the body center of mass in the x-axis direction, the minimum value is -0.2, the maximum value is 0.2, in meters (m); the body center of mass (y) is the position of the body center of mass in the y-axis direction, the minimum value is -0.15, the maximum value is 0.15, in meters (m); the minimum value of gripper load mass is 0.0 and the maximum value is 3.0, in kg; the minimum value of motor strength is The minimum value of proprioception delay is 0.005 and the maximum value is 0.045, and the unit is seconds (s); the minimum value of terrain height is 0.0 and the maximum value is 0.1, and the unit is m; the minimum value of gravity noise is -0.05 and the maximum value is 0.05; the minimum value of joint position noise is -0.01 and the maximum value is 0.01, and the unit is radians (rad); the minimum value of joint velocity noise is -1.5 and the maximum value is 1.5, and the unit is radians / second (rad / s); the minimum value of angular velocity noise is -0.2 and the maximum value is 0.2, and the unit is rad / s.

[0188] Table 1 Random setting parameters

[0189] parameter Minimum Maximum unit Ground friction 0.1 3.0 none Base load quality 0.0 20.0 kg Body center of mass (x) -0.2 0.2 m Body center of mass (y) -0.15 0.15 m Gripper load mass 0.0 2.0 kg Motor strength 70 130 ％ Delayed proprioception 0.005 0.045 s Terrain height 0.0 0.1 m Gravity Noise -0.05 0.05 none Joint position noise -0.01 0.01 rad Joint velocity noise -1.5 1.5 rad / s Angular velocity noise -0.20 0.20 rad / s

[0190] After training a robust teacher model, it is refined into a student model. The teacher model uses the privileged information encoder to generate a latent vector l t , and then generate action a through a multi-layer perceptron (MLP) t The student model replaces the encoder with a gated recurrent unit (GRU), which receives proprioceptive observation data o t prop And generate the latent vector l t , l t With o t prop After connection, MLP is used to generate action a t .

[0191] Yes t and a t For supervision, the definition of the loss function is shown in formula (1).

[0192] (2) Perception module;

[0193] The specific implementation process of the perception module is as follows: Fig.12 As shown, it may include steps S1201 to S1203:

[0194] Step S1201: ASGrasp grabs;

[0195] The infrared image is used as input to ASGrasp to predict the depth of the object.

[0196] Step S1202: image tracking;

[0197] Perform image segmentation and tracking on the color graphic image to obtain the image segmentation result.

[0198] Step S1203: Grab and select.

[0199] Generate real-time grasp poses through grasp selection.

[0200] In order to achieve real-time tracking and accurate grasping posture prediction, the disclosed embodiment adopts a video tracking and segmentation tool (Track Anything Model, TAM) and a generalized grasping technology (ASGrasp). After generating the initial mask using the image segmentation model (Segment Anything Model, SAM) in TAM, the memory module (XMem) outputs a real-time object mask based on the initial mask. Using infrared (IR) images and color images (RGB images) as input to ASGrasp, the accurate depth of transparent and reflective surfaces can be predicted. The generated depth point cloud is then input into the network model (Geometric and Scene-aware Network, GSNet) to generate a more accurate six-degree-of-freedom grasping posture.

[0201] (3) Operation module;

[0202] The specific implementation process of the operation module is as follows Fig.13 As shown, it may include steps S1301 to S1304:

[0203] Step S1301: The motion planner plans the trajectory of the robot arm;

[0204] The motion planner generates the robot arm trajectory based on the target grasping posture;

[0205] Step S1302: Determine the motor position using a differentiable kinematics model;

[0206] According to the trajectory of the robot arm, the motor position is determined through a differentiable kinematic model.

[0207] Here, the motor positions are the joint angles of the robot arm.

[0208] Step S1303: The PD controller determines the motor torque;

[0209] According to the motor position, the appropriate motor torque is determined by the PD controller.

[0210] Step S1304: Control the movement of the robotic arm.

[0211] Control the robot arm movement based on the motor torque.

[0212] The manipulation module is designed to actively track and grasp the target object while adjusting the basic motion of the robot. A motion planning approach is used to address the control error problem in the end-effector control using a full-body RL strategy. The system operates in two different phases: tracking and grasping.

[0213] During tracking, the movement of the camera mounted on the robot is constrained within a predefined tracking ball to reduce motion blur and prevent tracking loss. The tracking ball is parameterized by its position and radius in the fuselage frame. In the disclosed embodiment, a reachability map (RM) is used to define the tracking ball to ensure that the camera only works within the dexterous workspace, that is, a valid inverse kinematics (IK) solution can be obtained in any orientation. The switching mechanism is built based on the reachability map and the threshold criterion. In each planning step, the reachability of the selected grasping posture is calculated using RM. Once the threshold is reached, the system switches to the grasping phase. The motion planner can generate trajectories online, allowing the system to adapt to small unexpected movements while approaching the target.

[0214] (4) Planning module.

[0215] The specific implementation process of the planning module is as follows Fig.14 As shown, it may include steps S1401 to S1403:

[0216] Step S1401: determining a generalized directed reachability graph (GORM);

[0217] According to the real-time grasping posture, the generalized directional reachability map (GORM) is determined; according to the generalized directional reachability map, the reachability posture distribution of the fuselage can be determined.

[0218] Step S1402: Determine a planning strategy;

[0219] According to the real-time grasping posture, the planning strategy is determined and motion instructions are generated.

[0220] Step S1403: Determine the motion control strategy.

[0221] Generate motion control strategies based on motion instructions.

[0222] A reachability map (RM) is a posture quality metric that is often used to provide priors and targets for mobile manipulation tasks. An Oriented Reachability Map (ORM) can effectively represent potential basic poses (corresponding to the candidate base poses in the aforementioned embodiments) relative to a tool center point (TCP) frame. However, ORM is often applied to platform-based mobile manipulation, in which case the robot's base is limited to a plane. The disclosed embodiments propose a generalized Oriented Reachability Map (GORM), which supports robot base positioning in six degrees of freedom, such as Fig.15As shown, the generalized directional reachability graph 1502 is used to determine the candidate base position 1503 in the rectangular box, and the gripper grabs the object at the target position 1501.

[0223] For any target pose in the world frame (corresponding to the target gripper pose in the previous embodiment) p∈SE(3), the distribution of potential bases in the world coordinate system is calculated by inverse RM. The space is refined by removing the following basic poses: 1) out of range; 2) colliding with the environment; 3) below the reachability threshold. Once the target pose is defined, GORM provides a high-quality distribution of potential base positions. By training a high-level strategy, the distance between the current base pose and the nearest feasible position is minimized:

[0224] r GORM =exp(-(min(d GORM )) 2 ) (2);

[0225] At each high-level decision step, the distance distribution d is calculated by summing the Euclidean distance and the geodetic distance between the GORM and the current base pose. GORM , the minimum distance is selected from the distance distribution to encourage the robot to stay close to the nearest candidate base pose. Since GORM is defined in the target pose framework, it only needs to be calculated once, so it is very efficient and well suited for parallel training.

[0226] In vision-based control, it is crucial to train obstacle avoidance and object grasping strategies based on depth information. However, these methods usually introduce noise addition and filling techniques to bridge the gap between simulation and reality, but these techniques may reduce system performance in precision tasks such as grasping. In contrast, the method of the embodiment of the present disclosure does not rely on simulated depth images, but directly uses the grasping pose detection results from the off-the-shelf perception module.

[0227] The observation result definition of the advanced strategy is shown in formula (3):

[0228]

[0229] The vector Contains a rotation matrix rot t ∈R 9 and vector space mapping trans t ∈R 3 The gravity vector in the body frame is The joint position is q t ∈R 18 , and the last action is a t-1 ∈R 5 , which includes five-dimensional leg movement control instructions.

[0230] In the disclosed embodiment, a simulation experiment is first conducted on a whole-body grasping framework, in which the system is deployed on an onboard microcomputer.

[0231] Evaluation of the motion control module: The performance and energy efficiency of several motion controllers were compared, including the AsymAC controller, ROA controller, StateEstimator controller, and teacher-student controller. All methods were trained to track velocity instructions under consistent conditions and tested in a simulated environment with a large amount of randomization and observation noise to replicate real-world scenarios. Performance indicators were evaluated using instruction tracking rewards and average torque, respectively. With the goal of balancing low energy consumption and high performance, the teacher-student controller provided the best trade-off and was therefore selected as the core of the motion controller.

[0232] Evaluate the robot's entire working area: Fig.16 As shown, by tracking the instructions, the motion control strategy, under the guidance of the proprioceptive feedback of the arm, expands the whole body workspace without sacrificing robustness, and compares the size of the arm workspace 1601 and the whole body workspace 1602. Table 2 is an example of the two-dimensional area and three-dimensional volume of the arm workspace 1601 and the whole body workspace 1602. In Table 2, the two-dimensional area of ​​the arm workspace is 1.4, and the three-dimensional volume is 6.4; the two-dimensional area of ​​the whole body workspace is 2.15, and the three-dimensional volume is 8.54. It can be seen that the controller increases the effective workspace volume by 54% and the workspace area by 33%, and can effectively execute all instructions in various body postures (leaning forward, leaning back, squatting).

[0233] Table 2 Sizes of the arm workspace 1601 and the whole body workspace 1602

[0234] <![CDATA[Three-dimensional space volume (m 3 )]]> <![CDATA[Three-dimensional space area (m 2 )]]> Arm workspace 1.4 6.4 Full body workspace 2.15 8.54

[0235] Evaluation of general full-body grasping: The same simulation benchmark setting as VBC was used, that is, 34 objects were divided into 7 categories: ball, long box, square box, bottle, cup, bowl, drill. In the picking experiment, the position and orientation of the robot and the object were randomly reset at the beginning of each round. 300 trials were conducted for each object to evaluate the success rate, and the success condition was to pick up the target object before 150 high-level steps. The method of the embodiment of the present disclosure (i.e., the method with GORM reward), the method without GORM reward and the benchmark method were compared. For the method without GORM reward and the benchmark method, VBC was used to obtain the assistant reward. The purpose of this comparison is to verify whether the method of the embodiment of the present disclosure can accurately complete the mobile picking task of objects of different categories. As shown in Table 3, the success rate of picking up balls, long boxes, square boxes, bottles, cups, bowls, and drills by the method of the embodiment of the present disclosure, the method without GORM reward and the benchmark method, as well as the standard deviation of the success rate of grasping objects by these three methods. Among all test objects, the method of the embodiment of the present disclosure achieved a significantly higher success rate. It is worth noting that end-to-end RL methods such as VBC have a standard deviation of 14.53, reflecting considerable variability. VBC performs well on simple and small objects, such as square boxes (80%), but struggles on more complex or larger objects, such as drills (53.33%) and long boxes (28.57%). In contrast, the standard deviation of the method of the disclosed embodiment is much smaller, only 3.46, which shows that the performance of the method of the disclosed embodiment is more consistent on all object types. Regardless of the size or geometric complexity of the object, the method of the disclosed embodiment can show consistent performance. This robustness is largely due to the integration of the grasping posture detector, which provides less redundant input to the high-level strategy, allowing the system to adapt to different objects more effectively and ensure consistency of performance even in challenging situations. After the GORM reward was cancelled, the success rate dropped significantly. This drop may be due to the inability to guide the arm into the optimal workspace, especially in tasks that require precise manipulation skills.

[0236] Table 3 Success rate of grasping objects using the method with GORM reward, the method without GORM reward and the baseline method

[0237]

[0238] Evaluation of GORM-guided grasping: The same environment setup as above was used, but only the cup was used as the target object. Four different heights were set for the target object, namely 0m (floor), 0.3m (box), 0.75m (table), and 1m (shelf), and each was tested 500 times. After the 150th high-level step of each time, the average reachability of the target pose relative to the body frame was calculated. Three strategies were compared: the masked depth-based strategy, the strategy without GORM reward, and the strategy with GORM reward. The GORM reward used the grasp detection results obtained by the perception module, and for the masked depth-based strategy, the occluded depth image obtained from the camera sensor. The masked depth-based method and the method without GORM reward were trained according to the proximity reward and auxiliary reward designed in VBC. As shown in Table 5, the reachability results of the masked depth-based strategy, the strategy without GORM reward, and the strategy with GORM reward were tested for grasping objects on the floor, box, table, and shelf. The masked depth-based policy consistently underperforms at all height levels, struggling especially when the target position is below the default motion range of the base. Without the GORM reward, the policy suffers from reduced reachability at all heights as it is unable to find the optimal base grasp pose. In contrast, with the GORM reward, the policy maintains high reachability at all heights by guiding the motion module to find the optimal base position for grasping, maximizing performance regardless of the object height. As shown in Table 3, the overall grasp success rate is significantly improved with the GORM reward, suggesting that guiding the robot to the optimal base pose enables the manipulation module to perform more accurate motion planning, making the system more robust and reliable for grasping tasks.

[0239] In the disclosed embodiment, the full-body grasping framework is secondly evaluated in real events.

[0240] Comparison of VBC objects: The method of the disclosed embodiment was compared with the reinforcement learning method and the modular method. The former can grasp objects of different heights, while the latter can only effectively handle the setting of the table height. 14 real-world objects were collected and placed on the floor and the table in random postures. As shown in Table 4, the GAMMA method, the VBC method and the GROM reward method were used to perform grasping tests on objects on the floor, objects on the table, transparent objects and arbitrary objects, respectively, to obtain the grasping success rate results. The method of the disclosed embodiment performed best in all tasks. However, the GAMMA grasping method (i.e., the one-time grasping method) using a model-based low-level controller cannot grasp objects on the floor. VBC has a low success rate because it uses an RL-based control system.

[0241] Table 4 Comparison of the success rates of the GORM reward method, the GAMMA grasping method, and the baseline method under different object types and heights

[0242] floor desktop Transparent objects Any object GAMMA none 73％ none none VBC 30％ 12％ none none GROM rewards 90％ 75％ 80％ 89％

[0243] Table 5 Comparison of reachability of strategies based on occlusion depth, strategies without GORM reward, and strategies with GORM reward at different heights

[0244]

[0245] Experiment with arbitrary objects: Fig.17 The dataset used in the real-world test is presented, which includes objects from the YCB dataset and common recyclable items such as plastic bottles, cans, and crumpled paper or plastic packaging. In addition, the dataset also includes small toys that are often used in grasping tests, and a large part of the dataset is transparent objects of various shapes. During each trial, the dataset is shuffled and a set of objects are randomly dropped from a height of 1 meter to simulate random directions and positions. After simply marking the mask of the target object in the first frame, the automatic grasping process begins. Table 6 provides the results of 100 trials conducted in a real-world environment. In Table 6, the method of the embodiment of the present disclosure is used to perform the grasping object test, and the test results of the success rate, tracking error rate, grasping error rate, and object drop rate (the gripper falls after grasping the object) are obtained. It can be seen that the method of the embodiment of the present disclosure achieves a success rate of 89%. This proves the versatility of the whole-body grasping framework on a variety of objects, thanks to its modular design, which combines off-the-shelf perception models with powerful, learned motor skills.

[0246] Table 6 Results of 100 tests conducted in a real-world environment using the disclosed method

[0247] Success rate Tracking Error Rate Crawl Error Rate Drop rate 89.0％ 6％ 3％ 2％

[0248] Experiments on transparent objects: The disclosed embodiment conducted 10 experiments on grasping transparent objects, as shown in Table 6. The results show that in a cluttered scene, the grasping accuracy of a single camera reaches 80%. This challenging task is accomplished through adaptive coordination of various components and inherits the capabilities of the perception module without performance degradation.

[0249] Based on the above evaluation, the whole-body grasping framework combines the learned five-dimensional basic motion strategy with the generalized directional reachability graph (GORM) to achieve precise and robust positioning manipulation. This method effectively integrates precise manipulation with coordinated whole-body motion, and has significant improvements in handling various tasks in both simulated and real-world scenarios.

[0250] Based on the foregoing embodiments, the embodiments of the present disclosure provide a robot behavior planning device and a robot control method. The device includes the modules included, and the units included in the modules, etc., which can be implemented by a processor in a computer device; of course, it can also be implemented by a specific logic circuit; in the implementation process, the processor can be a central processing unit (CPU), a microprocessor (MPU), a digital signal processor (DSP) or a field programmable gate array (FPGA), etc.

[0251] Fig.18 A schematic diagram of the structure of a robot behavior planning device provided in an embodiment of the present disclosure is shown in FIG. Fig.18 As shown, the robot behavior planning device 1800 includes: a first determination module 1801, a second determination module 1802 and a third determination module 1803, wherein:

[0252] The first determination module 1801 is used to determine a target grasping posture of the robot grasping the target object based on the object posture of the target object; the target grasping posture includes a target base posture and a target gripper posture;

[0253] The second determination module 1802 is used to determine a motion control strategy of the robot's mechanical legs based on the target base seat posture and the current base seat posture of the robot; the motion control strategy is used to control the mechanical legs to drive the robot to move so that the base of the robot reaches the target base seat posture;

[0254] The third determination module 1803 is used to determine the grasping control strategy of the robot's mechanical arm based on the target grasper posture and the robot's current grasper posture; the grasping control strategy is used to control the mechanical arm to drive the robot's grasper to grasp the target object with the target grasper posture when the base reaches the target base posture.

[0255] In some embodiments, before determining the target grasping posture of the robot grasping the target object based on the object posture of the target object, the robot's behavior planning device includes: an acquisition unit for acquiring image information of the target object; and a first determination unit for identifying and locating the target object based on the image information to determine the object posture of the target object.

[0256] In some embodiments, the first determination unit includes: a first segmentation subunit, used to segment the target object in the color image to obtain a segmentation result; a first determination subunit, used to determine the depth of the target object based on the infrared image and the color image; and a second determination subunit, used to determine the object pose of the target object based on the segmentation result and the depth of the target object.

[0257] In some embodiments, the first determination module includes: a second determination unit, used to determine the target gripper posture of the gripper grasping the target object based on the object posture; and a third determination unit, used to determine the target base seat posture based on the target gripper posture.

[0258] In some embodiments, the third determination unit includes: a third determination subunit is used to determine, based on the target gripper posture, a plurality of candidate base seat postures so that the gripper can reach the target gripper posture; a fourth determination subunit is used to determine, based on the current base seat posture of the robot, the target base seat posture with the smallest distance from the current base seat posture from the plurality of candidate base seat postures.

[0259] In some embodiments, the second determination module includes: a fourth determination unit, used to determine the movement speed of the mechanical leg and the angles of each leg joint during the movement of the base from the current base posture to the target base posture based on the target base posture and the current base posture; a fifth determination unit, used to determine the motion control strategy based on the movement speed of the mechanical leg and the angles of each leg joint.

[0260] In some embodiments, the target base posture includes a target base position and a target base posture; the fourth determination unit includes: a fifth determination subunit, used to determine the movement speed of the mechanical legs and the leg joint angles during the movement of the base from the current base position to the target base position based on the current base position of the robot and the target base position; a sixth determination subunit, used to determine the leg joint angles of the mechanical legs during the movement of the base from the first posture to the target base posture based on the first posture of the base when the robot is at the target base position, and the target base posture.

[0261] In some embodiments, the target base posture includes a target base position and a target base posture; the fourth determination unit also includes: a seventh determination subunit, which is used to determine the movement speed of the mechanical leg and the angles of each leg joint in the process of the base reaching the target base position from the current base posture with the target base posture based on the current base posture and the target base posture.

[0262] In some embodiments, the second determination module includes: a first generation unit, used to generate a set of leg motion instructions for the mechanical leg based on the motion control strategy; a first control unit, used to control the mechanical leg to drive the robot to move based on the leg motion instruction set so that the base moves from the current base posture to the target base posture.

[0263] In some embodiments, the third determination module includes: a sixth determination unit, used to determine the motion trajectory of the robotic arm in the process of the gripper moving from the current gripper posture to the target gripper posture; a seventh determination unit, used to determine the arm joint angles of the robotic arm based on the motion trajectory; and an eighth determination unit, used to determine the grasping control strategy based on the arm joint angles of the robotic arm.

[0264] In some embodiments, the third determination module also includes: a second generation unit, used to generate an arm grasping instruction set for the robotic arm based on the grasping control strategy; a second control unit, used to control the robotic arm to drive the gripper from the current gripper posture to the target gripper posture to grasp the target object based on the arm grasping instruction set.

[0265] In some embodiments, the third determination module also includes: a third control unit, used to control the gripper to grasp the target object in the target gripper posture when the distance between the current base seat posture of the robot and the target base seat posture is less than or equal to a preset distance threshold.

[0266] Fig.19 A schematic diagram of the structure of a robot control device provided by an embodiment of the present disclosure is shown in FIG. Fig.19 As shown, the robot control device 1900 includes: a first acquisition module 1901, a second acquisition module 1902, a first control module 1903, a third acquisition module 1904 and a second control module 1905, wherein:

[0267] The first acquisition module 1901 is used to acquire the target grasping posture of the robot grasping the target object; the target grasping posture includes the target base posture and the target gripper posture;

[0268] A second acquisition module 1902 is used to acquire a motion control strategy of the robot's mechanical legs based on the target base posture;

[0269] A first control module 1903 is used to control the mechanical legs to drive the robot to move based on the motion control strategy, so that the base of the robot reaches the target base posture;

[0270] A third acquisition module 1904 is used to acquire a grasping control strategy of the robot's mechanical arm based on the target gripper posture;

[0271] The second control module 1905 is used to control the mechanical arm to drive the robot's gripper to grasp the target object in the target gripper posture when the base reaches the target base posture based on the gripping control strategy.

[0272] The description of the above device embodiment is similar to the description of the above method embodiment, and has similar beneficial effects as the method embodiment. In some embodiments, the functions or modules included in the device provided in the embodiments of the present disclosure can be used to execute the method described in the above method embodiment. For technical details not disclosed in the device embodiment of the present disclosure, please refer to the description of the method embodiment of the present disclosure for understanding.

[0273] It should be noted that in the embodiments of the present disclosure, if the above method is implemented in the form of a software function module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiment of the present disclosure is essentially or the part that contributes to the relevant technology can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the methods of each embodiment of the present disclosure. The aforementioned storage medium includes: various media that can store program codes, such as a U disk, a mobile hard disk, a read-only memory (ROM), a disk or an optical disk. In this way, the embodiments of the present disclosure are not limited to any specific hardware, software or firmware, or any combination of hardware, software, and firmware.

[0274] An embodiment of the present disclosure provides a computer device, including a memory and a processor, wherein the memory stores a computer program that can be run on the processor, and when the processor executes the program, some or all of the steps in the above method are implemented.

[0275] The embodiment of the present disclosure provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, some or all of the steps in the above method are implemented. The computer-readable storage medium can be transient or non-transient.

[0276] An embodiment of the present disclosure provides a computer program, including a computer-readable code. When the computer-readable code is run in a computer device, a processor in the computer device executes some or all of the steps for implementing the above method.

[0277] The present disclosure provides a computer program product, which includes a non-transitory computer-readable storage medium storing a computer program, and when the computer program is read and executed by a computer, some or all of the steps in the above method are implemented. The computer program product can be implemented in hardware, software, or a combination thereof. In some embodiments, the computer program product is specifically embodied as a computer storage medium, and in other embodiments, the computer program product is specifically embodied as a software product, such as a software development kit (SDK), etc.

[0278] It should be noted here that the description of the various embodiments above tends to emphasize the differences between the various embodiments, and the same or similar aspects can be referenced to each other. The description of the above device, storage medium, computer program and computer program product embodiments is similar to the description of the above method embodiment, and has similar beneficial effects as the method embodiment. For technical details not disclosed in the embodiments of the device, storage medium, computer program and computer program product disclosed in the present invention, please refer to the description of the method embodiment disclosed in the present invention for understanding.

[0279] It should be noted that Fig. 20 is a schematic diagram of a hardware entity of a computer device in an embodiment of the present disclosure, such as Fig. 20 As shown, the hardware entity of the computer device 2000 includes: a processor 2001, a communication interface 2002 and a memory 2003, wherein:

[0280] Processor 2001 generally controls the overall operation of computer device 2000 .

[0281] The communication interface 2002 enables the computer device to communicate with other terminals or servers through the network.

[0282] The memory 2003 is configured to store instructions and applications executable by the processor 2001, and can also cache data to be processed or processed by the processor 2001 and each module in the computer device 2000 (for example, image data, audio data, voice communication data, and video communication data), which can be implemented by flash memory (FLASH) or random access memory (Random Access Memory, RAM). Data can be transmitted between the processor 2001, the communication interface 2002, and the memory 2003 through the bus 2004.

[0283] It should be understood that "one embodiment" or "an embodiment" mentioned throughout the specification means that specific features, structures or characteristics related to the embodiment are included in at least one embodiment of the present disclosure. Therefore, "in one embodiment" or "in an embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. In addition, these specific features, structures or characteristics can be combined in one or more embodiments in any suitable manner. It should be understood that in the various embodiments of the present disclosure, the size of the serial numbers of the above-mentioned steps / processes does not mean the order of execution. The execution order of each step / process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present disclosure. The serial numbers of the embodiments of the present disclosure are for description only and do not represent the advantages and disadvantages of the embodiments.

[0284] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or device including the element.

[0285] In the several embodiments provided in the present disclosure, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are only schematic. For example, the division of units is only a logical function division. There may be other division methods in actual implementation, such as: multiple units or components can be combined, or can be integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the components shown or discussed can be through some interfaces, and the indirect coupling or communication connection of devices or units can be electrical, mechanical or other forms.

[0286] The units described above as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units; they may be located in one place or distributed on multiple network units; some or all of the units may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. In addition, the functional units in the various embodiments of the present disclosure may all be integrated into one processing unit, or each unit may be separately used as a unit, or two or more units may be integrated into one unit; the above integrated units may be implemented in the form of hardware or in the form of hardware plus software functional units.

[0287] A person skilled in the art can understand that all or part of the steps of implementing the above method embodiment can be completed by hardware related to program instructions, and the aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps of the above method embodiment; and the aforementioned storage medium includes: mobile storage devices, read-only memories (ROM), magnetic disks or optical disks, etc., various media that can store program codes.

[0288] Alternatively, if the above-mentioned integrated unit of the present disclosure is implemented in the form of a software function module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present disclosure can essentially or in other words, the part that contributes to the relevant technology can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the methods of each embodiment of the present disclosure. The aforementioned storage medium includes: various media that can store program codes, such as mobile storage devices, ROMs, magnetic disks, or optical disks.

[0289] The above description is only an implementation mode of the present disclosure, but the protection scope of the present disclosure is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed in the present disclosure should be included in the protection scope of the present disclosure.

Claims

1. A robot behavior planning method, characterized in that: The method comprises: Based on the object posture of the target object, determining the target grasping posture of the robot grasping the target object; the target grasping posture includes a target base posture and a target gripper posture; Determine a motion control strategy of the robot's mechanical legs based on the target base posture and the robot's current base posture; the motion control strategy is used to control the mechanical legs to drive the robot to move so that the robot's base reaches the target base posture; Based on the target gripper posture and the current gripper posture of the robot, a gripping control strategy of the robot's mechanical arm is determined; the gripping control strategy is used to control the mechanical arm to drive the robot's gripper to grip the target object at the target gripper posture when the base reaches the target base posture.

2. The method according to claim 1, characterized in that Before determining a target grasping posture of the robot grasping the target object based on the object posture of the target object, the method further includes: Acquiring image information of the target object; Based on the image information, the target object is identified and located, and the object posture of the target object is determined; The image information includes an infrared image and a color image; and based on the image information, identifying and locating the target object and determining the object position and posture of the target object include: Segmenting the target object in the color image to obtain a segmentation result; Determining the depth of the target object based on the infrared image and the color image; Based on the segmentation result and the depth of the target object, an object pose of the target object is determined.

3. The method according to claim 1, characterized in that The step of determining a target grasping posture of the robot grasping the target object based on the object posture of the target object comprises: Based on the object posture, determining a target gripper posture for the gripper to grasp the target object; Based on the target gripper posture, determining the target base seat posture; The step of determining the target base position based on the target gripper position comprises: Based on the target gripper posture, determining a plurality of candidate base seat postures for enabling the gripper to reach the target gripper posture; Based on the current base seat pose of the robot, a target base seat pose having the smallest distance to the current base seat pose is determined from a plurality of candidate base seat poses.

4. The method according to claim 1, characterized in that The step of determining a motion control strategy of a mechanical leg of the robot based on the target base seat posture and the current base seat posture of the robot comprises: Based on the target base posture and the current base posture, determining the movement speed of the mechanical leg and the angles of each leg joint during the movement of the base from the current base posture to the target base posture; The motion control strategy is determined based on the movement speed of the mechanical leg and the angles of each leg joint.

5. The method according to claim 4, characterized in that The target base posture includes a target base position and a target base posture; The step of determining the movement speed of the mechanical legs and the angles of each leg joint during the movement of the base from the current base posture to the target base posture based on the target base posture and the current base posture comprises at least one of the following: Based on the current base position of the robot and the target base position, determine the movement speed of the mechanical legs and the angles of each leg joint during the movement of the base from the current base position to the target base position; based on the first posture of the base when the robot is at the target base position and the target base posture, determine the angles of each leg joint of the mechanical legs during the movement of the base from the first posture to the target base posture; Based on the current base posture and the target base posture, the moving speed of the mechanical leg and the leg joint angles are determined in the process of the base reaching the target base position from the current base posture to the target base posture.

6. The method according to any one of claims 1 to 5, characterized in that The determining of the grasping control strategy of the robot's mechanical arm based on the target grasper posture and the robot's current grasper posture comprises: Determine the motion trajectory of the robotic arm during the movement of the gripper from the current gripper posture to the target gripper posture; Based on the motion trajectory, determining the angles of each arm joint of the robotic arm; The grasping control strategy is determined based on the arm joint angles of the robot arm.

7. The method according to any one of claims 1 to 5, characterized in that When the base reaches the target base posture, controlling the mechanical arm to drive the gripper of the robot to grasp the target object in the target gripper posture includes: When the distance between the current base seat posture and the target base seat posture is less than or equal to a preset distance threshold, the gripper is controlled to grip the target object with the target gripper posture.

8. A robot control method, characterized in that: The method comprises: Acquire a target grasping posture of the robot grasping the target object; the target grasping posture includes a target base posture and a target gripper posture; Acquire a motion control strategy of the robot's mechanical legs based on the target base posture; Based on the motion control strategy, controlling the mechanical legs to drive the robot to move so that the base of the robot reaches the target base posture; Acquire a grasping control strategy of the robot's mechanical arm based on the target gripper posture; Based on the grasping control strategy, when the base reaches the target base posture, the mechanical arm is controlled to drive the gripper of the robot to grasp the target object in the target gripper posture.

9. A robot comprising a planning module, a base, a gripper, a mechanical arm and a mechanical leg, wherein: The planning module is used to determine the target grasping posture of the robot for grasping the target object based on the object posture of the target object; the target grasping posture includes a target base posture and a target gripper posture; based on the target base posture and the current base posture of the robot, the motion control strategy of the robot's mechanical legs is determined; the motion control strategy is used to control the mechanical legs to drive the robot to move so that the base of the robot reaches the target base posture; based on the target gripper posture and the current gripper posture of the robot, the grasping control strategy of the robot's mechanical arm is determined; the grasping control strategy is used to control the mechanical arm to drive the robot's gripper to grasp the target object in the target gripper posture when the base reaches the target base posture.

10. The robot according to claim 9, characterized in that: The robot further includes at least one of a perception module, a motion control module, an operation module, and an image acquisition component, wherein: The perception module is used to obtain image information of the target object; based on the image information, identify and locate the target object to determine the object position and posture of the target object; Based on the motion control strategy, a leg motion instruction set of the mechanical leg is generated; based on the leg motion instruction set, the mechanical leg is controlled to drive the robot to move so that the base moves from the current base posture to the target base posture; The operation module is used to generate an arm grasping instruction set of the robot arm based on the grasping control strategy; based on the arm grasping instruction set, control the robot arm to drive the gripper to move from the current gripper posture to the target gripper posture to grasp the target object; The image acquisition component is arranged on the mechanical arm; the image acquisition component is used to acquire image information of the target object.

Images