Whole-body teleoperation method for humanoid robot based on joint matching and intention recognition

Abstract

The application discloses a kind of whole body teleoperation methods of humanoid robot based on joint matching and intention recognition, comprising: collecting the first attitude data of operator upper body joint and the second attitude data of lower limb joint;First attitude data is mapped to each corresponding upper body joint of robot;Based on second attitude data, motion intention is extracted by hybrid hierarchical logic, and lower limb intention instruction is obtained;Based on lower limb intention instruction, the course reinforcement learning framework based on logarithmic attenuation mechanism is used as control strategy, and the control torque of each lower limb joint of robot is generated, to realize the lower limb teleoperation control of robot.The application is applied to the field of robot control, not only can completely liberate the hands and waist of operator, so that it obtains more teleoperation degrees of freedom, but also can make robot stable in actual teleoperation, avoid instability when facing large amplitude and rapid arm swing and other violent upper body inertia disturbance.

Description

Technical Field

[0001] This invention relates to the field of robot control technology, specifically a method for whole-body teleoperation of a humanoid robot based on joint matching and intent recognition. Background Technology

[0002] With the development of humanoid robot teleoperation technology, in scenarios such as human-robot collaborative work and complex environment operations, operators typically need to complete tasks using their upper bodies, while the robot's lower limbs achieve autonomous and stable movement. Full-body teleoperation of humanoid robots requires the end-user robot to possess extremely strong anti-disturbance balance capabilities to cope with the inertial impact generated by unpredictable large-amplitude human movements. However, existing linear learning degenerates into uniformly distributed sampling in the later stages of training, resulting in insufficient exploration of extreme boundary conditions (such as large-amplitude, rapid arm swings). When severe upper-body inertial disturbances occur during actual teleoperation, the robot is highly susceptible to instability.

[0003] In traditional reinforcement learning (RL) curriculum design, the evolution of the probability density function for disturbance terms (such as the amplitude of violent upper body movements) is not flexible enough, and linear decay is usually used. This linear curriculum has problems with absolute safety in the early stage of training and difficulty in hyperparameter adjustment for effective sampling and targeted training of long-tailed distributions (extreme, violent movements) in the later stage of training. It is difficult to ensure smooth convergence of training while allowing the policy to fully learn the balance ability in extreme dynamic scenarios. Ultimately, this results in insufficient robustness and poor stability of the robot when faced with the high dynamic upper body movements of the operator in real teleoperation. Summary of the Invention

[0004] To address the shortcomings of the existing technologies, this invention provides a method for full-body teleoperation of a humanoid robot based on joint matching and intent recognition. This method not only completely frees the operator's hands and waist, giving them more degrees of freedom in teleoperation, but also enables the robot to maintain stability and avoid instability when faced with severe upper-body inertial disturbances such as large-amplitude, rapid arm swings during actual teleoperation.

[0005] To achieve the above objectives, the present invention provides a method for whole-body teleoperation of a humanoid robot based on joint matching and intent recognition, comprising the following steps: Data acquisition: The operator wears a full-body exoskeleton device, which includes an upper body exoskeleton and a lower limb exoskeleton. The upper body exoskeleton and the lower limb exoskeleton respectively collect the first posture data of the operator's upper body joints and the second posture data of the lower limb joints. Upper body joint mapping: The first posture data is mapped to the corresponding upper body joints of the robot to realize the upper body remote operation control of the robot; Lower limb intention recognition: Based on the second posture data, the motion intention is extracted through hybrid hierarchical logic to obtain the lower limb intention command; Lower limb joint control: Based on the lower limb intention command, a course reinforcement learning framework based on logarithmic decay mechanism is used as the control strategy to generate control torques for each lower limb joint of the robot, thereby realizing the lower limb teleoperation control of the robot.

[0006] Compared with the prior art, the present invention has the following beneficial technical effects: 1. This invention uses a hybrid layered logic to extract movement intentions by collecting only the operator's lower limb posture data, without relying on upper body posture data to control the lower limbs. This completely liberates the operator's hands and waist, giving them more freedom of teleoperation and greatly improving the flexibility and convenience of teleoperation. 2. This invention employs a logarithmic decay mechanism-based curriculum reinforcement learning framework to replace the traditional linear decay curriculum. This ensures absolute safety in the early stages of training (concentrated small-amplitude perturbation sampling to ensure smooth policy convergence) while enabling effective sampling and targeted training of long-tailed distributions (extreme, intense upper body movements) in the later stages of training. It solves the problem that existing linear curricula cannot effectively balance the relationship between the early "cold start" and the later exploration. As a result, when the robot faces intense upper body perturbations in actual teleoperation, the fluctuation area of ​​the plantar pressure center (CoP) is significantly reduced, effectively avoiding oscillations, controlling the convergence speed faster, and greatly improving the lower limb balance stability. This allows the robot to adapt to highly dynamic upper body movements in complex teleoperation scenarios. Attached Figure Description

[0007] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0008] Figure 1 This is a flowchart of a whole-body teleoperation method for a humanoid robot based on joint matching and intent recognition, as described in an embodiment of the present invention. Figure 2 This is a first axonometric view of the full-body exoskeleton device in an embodiment of the present invention; Figure 3 This is a second axonometric view of the full-body exoskeleton device in an embodiment of the present invention; Figure 4 This is a third axonometric view of the full-body exoskeleton device in an embodiment of the present invention; Figure 5 This is a partial schematic diagram of the back support unit and the waist connection unit in an embodiment of the present invention; Figure 6 This is a schematic diagram of the hip joint drive unit in an embodiment of the present invention; Figure 7 This is a front view of the teleoperation synchronization effect in an embodiment of the present invention; Figure 8 This is a rear view of the teleoperation synchronization effect in an embodiment of the present invention.

[0009] Reference numerals: Upper body unit 1, Back support unit 2, Shoulder strap plate 21, Buckle base fastener 211, Buckle base 212, Buckle No. 1 213, Buckle No. 2 214, Back inertial measurement unit No. 1 215, Back inertial measurement unit No. 2 216, Waist connection unit 3, Waist main support 31, Waist belt 311, Right thigh support 32, Left thigh support 33, Hip joint drive unit 4, Right hip joint drive motor 41, First output flange 411, Left hip joint drive motor 42 421 Second output flange, 43 Right thigh binding, 44 Left thigh binding, 45 Thigh joint inertial measurement unit 1, 47 Right calf support, 48 Left calf support, 5 Knee joint drive unit, 51 Right knee joint drive motor, 511 Third output flange, 52 Left knee joint drive motor, 521 Fourth output flange, 53 Right calf binding, 54 Left calf binding, 55 Calf joint inertial measurement unit 1, 56 Calf joint inertial measurement unit 2, 6 Dexterous hand unit.

[0010] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0011] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0012] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indication will also change accordingly.

[0013] Furthermore, in this invention, descriptions involving "first," "second," etc., are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0014] In this invention, unless otherwise explicitly specified and limited, the terms "connection," "fixed," etc., should be interpreted broadly. For example, "fixed" can mean a fixed connection, a detachable connection, or an integral part; it can mean a mechanical connection, an electrical connection, a physical connection, or a wireless communication connection; it can mean a direct connection or an indirect connection through an intermediate medium; it can mean the internal communication of two elements or the interaction between two elements, unless otherwise explicitly limited. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0015] Furthermore, the technical solutions of the various embodiments of the present invention can be combined with each other, but only if they are feasible for those skilled in the art. If the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0016] This embodiment discloses a whole-body telemanipulation method for a humanoid robot based on joint matching and intent recognition, mainly comprising four parts: data acquisition, upper body full-degree-of-freedom exoskeleton joint mapping, lower limb multimodal exoskeleton intent recognition, and lower limb motion control strategy. The data acquisition part is implemented through a whole-body exoskeleton device. Upper body full-degree-of-freedom exoskeleton joint mapping relies on joint motors to map the shoulder and elbow, mapping 3 degrees of freedom for the shoulder and 1 degree of freedom for the elbow, respectively. For weight reduction, the mapping of the waist and wrist is obtained through the relative rotation of inertial sensing units (IMUs). The upper body directly controls the robot's upper-body movement through joint mapping. Lower limb multimodal exoskeleton intent recognition uses information acquired from the IMU to identify the intent of human movement and uses it as control commands for the lower limb motion control strategy. The lower limb motion control strategy is implemented using an actor-critic reinforcement learning network based on the proximal policy optimization (PPO) algorithm. An upper body interference course mechanism is introduced during the training process, allowing the robot to undergo a progressive learning process in the training environment, from cold start with a tendency to a safe position, to uniform sampling, and then to extreme sampling in a wide range of extreme positions. This enables the robot to maintain teleoperation balance accurately and stably.

[0017] refer to Figure 1 The whole-body teleoperation method for a humanoid robot based on joint matching and intent recognition in this embodiment specifically includes the following steps: Data acquisition: The operator wears a full-body exoskeleton device, which includes an upper body exoskeleton and a lower limb exoskeleton. The upper body exoskeleton and the lower limb exoskeleton respectively collect the first posture data of the operator's upper body joints and the second posture data of the lower limb joints. Upper body joint mapping: Mapping the first posture data to the corresponding upper body joints of the robot to realize the remote operation control of the robot's upper body; Lower limb intention recognition: Based on the second posture data, the movement intention is extracted through hybrid hierarchical logic to obtain the lower limb intention command; Lower limb joint control: Based on lower limb intention commands, a course reinforcement learning framework with random upper body joint targets as dynamic perturbations and a logarithmic decay mechanism is adopted as the control strategy to generate control torques for each lower limb joint of the robot, thereby realizing the robot's lower limb teleoperation control.

[0018] In this embodiment, the full-body exoskeleton device is the core hardware for teleoperation and imitation learning training of the humanoid robot. The operator wears the full-body exoskeleton device to synchronously drive the full-body movements of the humanoid robot. (Reference) Figures 2 to 4 The full-body exoskeleton device includes an upper body exoskeleton and a lower limb exoskeleton. The upper body exoskeleton includes an upper body unit 1, a back support unit 2, a waist connection unit 3, and a dexterous hand unit 6. The lower limb exoskeleton includes a hip joint drive unit 4 and a knee joint drive unit 5.

[0019] The upper body unit 1 is a completely symmetrical dual-arm structure. Each side of the upper body exoskeleton includes a shoulder coupling drive module, an upper arm support module, an elbow dual-degree-of-freedom drive module, and a forearm support module. The shoulder coupling drive module is hinged at its root to the shoulder mounting position of the back support unit 2's shoulder strap plate via crossed roller bearings. It employs a linkage-synchronous belt hybrid transmission structure, adapted to the human shoulder-humeral rhythm, and uses a servo drive motor with an absolute encoder to achieve high-precision tracking of shoulder compound movements. Six-dimensional force sensors are connected in series between the upper arm support module and the elbow drive module, and between the forearm support module and the elbow drive module, enabling real-time acquisition of human-machine interaction force data and achieving bidirectional force feedback and motion data acquisition. The forearm support module has a quick-release docking interface at its end for rapid mechanical and electrical connection with the dexterous hand unit. The single-sided upper body exoskeleton fully covers all seven degrees of freedom of movement of one side of the humanoid robot's upper body, achieving a 1:1 precise mapping between human upper body movements and humanoid robot upper body movements, driving the robot's upper body to perform synchronized movements with the operator wearing the full-body exoskeleton.

[0020] The dexterous hand unit 6 also features a completely symmetrical two-handed structure, including a right-hand dexterous hand data glove and a left-hand dexterous hand data glove. Each dexterous hand data glove comprises the glove body and a wrist IMU. The dexterous hand data gloves can fit snugly against the operator's hand, adapting to operators with different hand shapes. The wrist IMU is fixed to the back of the glove's wrist and completes electrical and mechanical docking with the forearm support module of the upper body exoskeleton unit through a quick-release docking interface. It can collect posture data of the wrist's three rotational degrees of freedom in real time, which is used for remote operation control and imitation learning training data acquisition of the humanoid robot's dexterous hand, driving the robot's hand grasping and manipulating actions to be synchronously reproduced by the operator when wearing the full-body exoskeleton device.

[0021] In the specific implementation process, the upper body unit 1 and the dexterous hand unit 6 can directly adopt the existing technology, the structure disclosed in Chinese invention patent application, publication number CN119427324A, which will not be described in detail in this embodiment.

[0022] refer to Figure 5 The back support unit 2 includes a shoulder strap plate 21, a buckle fixing assembly, and a dual inertial measurement unit. The buckle fixing assembly includes a buckle base fixing member 211 fixed to the back of the shoulder strap plate 21, a buckle base 212 locked onto the buckle base fixing member 211, and buckles 213 and 214 provided on the buckle base 212. The dual inertial measurement unit includes a back inertial measurement unit 1 215 and a back inertial measurement unit 216. The back inertial measurement unit 1 215 and the back inertial measurement unit 216 are detachably fixed to the shoulder strap plate 21 by buckles 1 213 and 214, respectively.

[0023] refer to Figures 2 to 5 The waist connection unit 3 includes a waist main support 31, on which a waist belt 311 for wearing and fixing is provided. The upper end of the waist main support 31 is separated from the lower end of the back strap plate 21. The right thigh support 32 is hinged to the right side of the waist main support 31, and the left thigh support 33 is hinged to the left side. The hinge axis is set along the front-back direction of the human body, realizing the lateral swing freedom of the lower limbs relative to the torso. It can flexibly adapt to the operator's stride width adjustment and lateral movement. The collected waist posture data is synchronously mapped to the waist joint of the slave humanoid robot, realizing the synchronous following of the robot's waist movement and the operator's torso posture. The rear side of the waist main support 31 is provided with a slot that matches the No. 2 buckle 214 for fixing the No. 2 back IMU 216 after calibration.

[0024] For upper body teleoperation, the humanoid robot's shoulders can be teleoperated via the scapula and shoulder motors of the exoskeleton, and the elbows can be teleoperated via elbow motor mapping. Wrist and waist teleoperation is calculated through the relative rotation of paired inertial sensors. The human waist and wrists are modeled as ball joints, and the inertial measurement unit quaternion of the reference trunk / forearm is represented as follows: The quaternion representation of the inertial measurement unit at the waist / wrist is as follows: The rotation quaternion for the transformation from the reference torso / forearm IMU coordinate system to the waist / wrist IMU coordinate system is:

[0025] in, Quaternion The conjugate is on the exoskeleton; during initialization or when lumbar / wrist calibration is required, the operator returns the lumbar / wrist IMU to its initial position and records the IMU readings at this time. and The quaternion calculation is as follows:

[0026] Therefore, the posture quaternion of the waist / wrist relative to the reference torso / forearm can be obtained at any time as:

[0027] This quaternion will serve as the target pose input for the waist / wrist, and through the iterative inverse kinematics (IK) method, the pose will be mapped onto the three joints of the robot's waist / wrist, thereby enabling teleoperation of the waist / wrist.

[0028] refer to Figures 2 to 4 The hip joint drive unit 4 includes a right hip joint drive motor 41, a left hip joint drive motor 42, a right thigh binding 43, a left thigh binding 44, a thigh posture sensing unit, and a lower leg connecting bracket. The stator end of the right hip joint drive motor 41 is fixedly connected to the right thigh bracket 32, and its rotor end is coaxially fixed with a first output flange 411. The first output flange 411 is locked and fixed to the right thigh binding 43, and the pitching motion of the thigh relative to the torso is directly achieved through the rotation drive of the motor. The stator end of the left hip joint drive motor 42 is fixedly connected to the left thigh bracket 33, and its rotor end is coaxially fixed with a second output flange 421. The second output flange 421 is locked and fixed to the left thigh binding 44, and the pitching drive of the left thigh is achieved synchronously. The structures on the left and right sides are completely symmetrical.

[0029] refer to Figure 6The thigh posture sensing unit includes thigh joint inertial measurement unit 1 (45) and thigh joint inertial measurement unit 2, which are fixed to the side mounting positions of the right thigh binding 43 and left thigh binding 44, respectively. These units are used to collect angular velocity and acceleration data of the thigh segment in real time, sensing the motion state and spatial posture of the lower limb. The lower leg connecting bracket includes a right lower leg bracket (47), fixed to the lower end of the right thigh binding 43, for detachable connection to the knee joint drive unit. The lower limb motion data collected by the thigh posture sensing unit is synchronized to the main control box in real time, mapped to control commands for the hip joint of the slave-end humanoid robot, driving the robot's hip joint and the operator's thigh to complete synchronized pitch and lateral movements. The left thigh posture sensing unit and left lower leg bracket (48) are also included in this configuration, and will not be described further in this embodiment.

[0030] The knee joint drive unit 5 is a detachable modular structure, including a right knee joint drive motor 51, a left knee joint drive motor 52, a right lower leg restraint 53, a left lower leg restraint 54, and a lower leg posture sensing unit. The stator end of the right knee joint drive motor 51 is hinged to the right lower leg support 47 to achieve the adaptation of the lateral swing freedom of the lower leg. Its rotor end is coaxially fixed with a third output flange 511, which is locked and fixed to the right lower leg restraint 53. The flexion and extension movement of the lower leg relative to the thigh is achieved by the rotation drive of the motor. The stator end of the left knee joint drive motor 52 is hinged to the left lower leg support 48. Its rotor end is coaxially fixed with a fourth output flange 521, which is locked and fixed to the left lower leg restraint 54, synchronously achieving the flexion and extension drive of the left lower leg. The structures on the left and right sides are completely symmetrical.

[0031] The lower leg posture sensing unit includes a lower leg joint inertial measurement unit 1 (55) and a lower leg joint inertial measurement unit 2 (56), which are fixed to the right lower leg binding 53 and the left lower leg binding 54, respectively. These units are used to collect real-time motion posture data of the lower leg segment, providing closed-loop feedback for hip-knee joint motion control. The collected lower leg motion data is synchronously mapped to control commands for the humanoid robot's knee joint, driving the robot's knee joint and the operator's lower leg to complete synchronized flexion and extension movements. This can completely reproduce the operator's lower limb movement postures such as squatting, standing up, and walking.

[0032] In this embodiment, the calibration process for the No. 1 back inertial measurement unit 215 and the No. 2 back inertial measurement unit 216 is as follows: During actual wear and use, the No. 1 back inertial measurement unit 215 is first fixed to the mounting position of the back strap plate 21 using buckle 1 213, and the No. 2 back inertial measurement unit 216 is fixed to the same reference zeroing position of the back strap plate 21 using buckle 214. With both inertial measurement units on the same rigid reference plane, coordinate system zeroing is performed on both inertial measurement units to ensure that their relative position references are unified. Then, keeping the No. 1 back inertial measurement unit 215 fixed in the mounting position of the back strap plate 21, the No. 2 back inertial measurement unit 216, along with buckle 214, is removed from the back strap plate 21 and snapped into the preset slot on the back of the waist main support 31. This quickly completes the determination of the relative position of the upper body and lower limbs, greatly simplifying the calibration process for coordinated control of the upper and lower limbs.

[0033] like Figure 7 , Figure 8 As shown, after entering the remote operation synchronous drive mode, when the operator wearing the full-body exoskeleton performs a horizontal abduction movement of both arms, the shoulder and elbow encoders of the upper body exoskeleton collect joint rotation angle data in real time. After calculation, the data is sent to the slave humanoid robot in real time, and the robot's arms synchronously complete the abduction movement at the same angle as the wearer. When the operator performs a squatting movement with knee flexion, the inertial measurement units of the hip joint drive unit and the knee joint drive unit collect the flexion and extension angle data of the hip and knee joints in real time. The main control box maps these data into control commands for the robot's hip and knee joints, and the robot synchronously performs the knee flexion movement with the same squat depth, achieving synchronous following of the whole body posture.

[0034] In this embodiment, the motion control of the humanoid robot, such as walking and turning, is not achieved through joint mapping. Instead, the operator's intention is extracted, and the operator's lower limbs are controlled by sending commands. The extraction of lower limb intention commands in this embodiment is implemented through hybrid hierarchical logic, which mainly includes three parts: lower limb motion data acquisition, motion feature extraction, and intention recognition and command generation based on hybrid hierarchical logic.

[0035] Lower limb motion data acquisition utilizes four inertial measurement units on the lower limb exoskeleton to collect leg pitch angle data in real time, i.e., the aforementioned second posture data, including the left thigh pitch angle. Right thigh pitch angle left lower leg pitch angle and the right lower leg tilt angle .

[0036] Motion feature extraction is based on the pitch angle of the left thigh. Right thigh pitch angle left lower leg pitch angle and the right lower leg tilt angle Extracting thigh symmetry features Symmetrical features of the lower legs Left leg flexion characteristics Characteristics of right leg flexion The extraction process is as follows:

[0037]

[0038]

[0039]

[0040] Among them, thigh symmetry features Symmetry features of the lower leg Used to characterize the operator's lower limb motor coordination, left leg flexion characteristics Characteristics of right leg flexion Used to quantify the degree of flexion of the operator's limbs.

[0041] After feature extraction is completed, the extracted features are input into the hybrid hierarchical logic, which includes a symmetry judgment layer and a dual-branch recognition layer.

[0042] The symmetry judgment layer is based on fuzzy set theory and evaluates the symmetry of the thigh. Symmetry with calves To distinguish between symmetrical squatting and asymmetrical walking / turning. To overcome random uncertainties and measurement noise in human dynamic perception, this embodiment uses a Gaussian membership function. Modeling (where) The mean, (Standard deviation), meaning the system calculates the trigger strength of each rule based on the set fuzzy logic. And combined with the corresponding symmetric confidence weights Calculate and output the final symmetry score. ,for:

[0043] in, The constant is set to a very small value to prevent the denominator from being zero. As for setting fuzzy logic and calculating the trigger strength, these are conventional techniques in the field, and will not be described in detail in this embodiment.

[0044] The dual-branch recognition layer determines the symmetry score based on the output of the symmetry judgment layer. With the preset symmetry threshold Activate the corresponding control branch to generate robot control instructions: Branch 1 (Symmetric Branch / Squat Continuous Mapping): When symmetry scores Greater than the symmetry threshold If the system detects a squatting motion, it determines the current movement state as "squatting." At this point, the system will calculate the average knee flexion of both legs. As input, shallow, medium, and deep fuzzy Gaussian membership degrees (denoted as ) are used. Outputs continuous squat depth commands. ,for:

[0045] After receiving the squat depth command Then, it can be output as a lower limb intention command; Branch 2 (Asymmetric Branch / Discrete Motion Rule Base): When symmetry score Less than or equal to the symmetry threshold When the current motion state is determined to be "walking" or "turning", the branch bypass fuzzy estimation adopts a deterministic rule logic module, that is, it uses a preset reference angle. For example, setting Discretize the real-time posture of limb segments into "upright" (greater than) ) or "tilt" (less than or equal to The system matches the tilt combination states of each leg segment with asymmetric logic rules, thereby directly triggering and sending walking or turning commands to the humanoid robot. The specific rule base is shown in Table 1.

[0046] Table 1 Asymmetric Branching Rule Base

[0047] For the logical rules in Table 1, the output motion intent label is specifically defined as: Walking: Turning: Among them, symbols Used to indicate directional information, this directional feature is determined by the side of the limb in a "tilted" state. For example, tilting the left limb will trigger a positive command, while tilting the right limb will trigger a negative command.

[0048] Therefore, in this embodiment, the process of obtaining the lower limb intention command under branch two is as follows: First, preset the reference angle. Adjust the pitch angle of the left thigh. The pitch angle of the right thigh The pitch angle of the left lower leg And the pitch angle of the right lower leg Respectively with reference angle Compare, and select those that are larger than the reference angle. The marked one is in an upright position, and the other marked ones are in a tilted position; When the supporting thigh and lower leg are both upright and the non-supporting thigh and lower leg are both tilted, it is determined to be a walking movement, and the obtained walking command is output as a lower limb intention command. When both the left and right thighs are upright, and one of the left and right lower legs is upright while the other is tilted, it is determined to be a turning movement, and the resulting turning command is output as a lower limb intention command.

[0049] After receiving the intention command from the lower limbs, the control strategy can be used to generate the control torque of each lower limb joint of the robot, realizing the remote operation control of the robot's lower limbs. This eliminates the need for the system to rely on traditional handheld joysticks or fixed pedals. Wearable devices replace the physical control panel, completely freeing the operator's hands and releasing the degree of freedom of movement of the waist. Ultimately, this achieves unconstrained full-body coordinated remote operation of the humanoid robot.

[0050] In this embodiment, the control system corresponding to the control strategy adopts an actor-critic network architecture and is trained based on a near-end policy optimization algorithm. The control system at each control step... The received observation state vector is ,in, For lower limb intention commands, including the desired horizontal linear velocity in the trunk coordinate system. , and the desired yaw rate and expected torso height . , These represent the base angular velocity and gravity vector of the robot in the torso coordinate system. , These represent the joint positions and velocity vectors of the entire robot. This represents the robot's lower limb motion vector from the previous step; To capture the time dependence of motion, the input to the control system at the current control step is obtained by concatenating the observation vectors of the current step and the past 5 steps. That is, at the current control step, the control system uses the concatenated input vector... This yields the lower limb motion vector for the current step. ; Based on lower limb motion vectors Calculate the target joint position of the robot's lower limb ,for:

[0051] in, This represents the natural configuration of the robot's lower limb joints. for 3D real space, This refers to the number of joints in the robot's lower limbs. In contrast, the upper body joints are directly controlled remotely by the operator and are therefore not included in the motion vectors. middle; After obtaining the location of the target joint in the lower limb Then, the control torques for each lower limb joint of the robot are generated by the PD controller, as follows:

[0052] in, For the robot Control torque of each lower limb joint , The robots are respectively Proportional gain and differential gain of each lower limb joint , The robots are respectively The actual position and actual speed of each lower limb joint.

[0053] To ensure the stability of teleoperation, this invention proposes a hierarchical curriculum learning strategy to adjust the upper body range of motion, thereby serving as a perturbation during robot training. A curriculum reinforcement learning framework based on a logarithmic decay mechanism is employed during the training of the control system corresponding to the control strategy. Specifically, this includes: Introducing a difficulty factor By dynamically adjusting the amplitude boundary of random upper body perturbations, when the control strategy exhibits sufficient motion stability, this difficulty factor dynamically modulates the random distribution of upper body joint targets and updates it based on the average reward to adjust the training difficulty of random upper body joint movements. Motion Amplitude Boundary The sampling follows a truncated exponential distribution. This embodiment proposes a logarithmic decay parameter. To control the evolution of the distribution, i.e., to calculate the logarithmic decay parameter based on the difficulty factor. And based on the logarithmic decay parameter Constructing the truncated exponential sampling distribution ,for:

[0054]

[0055] in, The initial attenuation coefficient, This is the scaling factor, which can be set to [value] in the specific implementation process. This design achieves a smooth transition: from concentrated, minute movements that ensure stability. Transition to uniform sampling Ultimately, it leans towards the kinematic limit. To master an extreme stance; Finally, based on the truncated exponential sampling distribution Perform inverse transformation sampling to generate corresponding random upper body joint target positions. This approach aims to achieve progressive training, focusing on small-amplitude perturbations in the early stages and shifting towards kinematically extreme perturbations in the later stages, with target positions for the upper body joints. Specifically:

[0056] in, It is a standard uniform random variable; Based on the aforementioned logarithmic decay mechanism, a composite reward function is constructed that includes speed tracking reward and composite height reward to guide the control strategy to achieve lower limb stability control under upper body dynamic disturbances.

[0057] In this embodiment, the speed tracking reward is specifically as follows:

[0058] in, For speed tracking rewards, For the command velocity vector, This is the actual velocity vector. Sensitivity coefficient; For vertical control, adjusting the base height is crucial for expanding the feasible operating space. To overcome the balance problem caused by the shift in the center of gravity during squats, this embodiment designs a composite height reward. To minimize height deviation while forcing the robot to lean backward during a squat, thus avoiding the problem of forward pacing when height changes rapidly, the following steps are taken:

[0059] in, For compound height rewards, For the desired torso height, This refers to the actual height of the base relative to the robot's feet. The attitude term weight (preferably set to 0.2) For squat range indicator, The gravity vector For compensation coefficient, The sensitivity factor is [not specified]. The squat range indicator only activates when the command falls within the squat range ([not specified]). Activated when ). Adaptive goals ( )(in Guide the projection of the gravity vector Reaching a negative value proportional to the squat depth effectively shifts the torso backward to counteract the tendency to lean forward. This strategy effectively prevents the robot from performing compensatory movements to maintain balance.

[0060] In the specific implementation process, the update process of the difficulty factor is as follows:

[0061] in, For the preset threshold, The average reward refers to the average of all rewards, including speed tracking rewards, composite height rewards, and other rewards.

[0062] The humanoid robot full-body teleoperation method based on joint matching and intent recognition in this embodiment generates fused control commands by accurately mapping the inverse kinematics joints of the upper body and recognizing the motion intent of the lower limbs based on hybrid hierarchical logic. This breaks the spatial limitations of traditional teleoperation and achieves unconstrained full-body teleoperation control of the slave humanoid robot. Simultaneously, the underlying balance and motion control strategies of the slave robot are globally allocated by a curriculum reinforcement learning framework based on a logarithmic decay mechanism. This framework treats the operator's upper body movements as controlled disturbances and continuously evolves through composite rewards and symmetry constraints, ultimately achieving extremely stable and efficient human-robot collaborative execution within a complex workspace.

[0063] It is worth noting that, although this embodiment Figure 1 The steps are shown sequentially as indicated by the arrows, but they are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order in which these steps are performed; they can be executed in other orders. Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0064] The above description is only a preferred embodiment of the present invention and does not limit the scope of protection of the present invention. All equivalent structural transformations made under the inventive concept of the present invention using the contents of the present invention specification and drawings, or direct / indirect applications in other related technical fields, are included within the scope of protection of the present invention.

Claims

1. A method for whole-body teleoperation of a humanoid robot based on joint matching and intent recognition, characterized in that, Includes the following steps: Data acquisition: The operator wears a full-body exoskeleton device, which includes an upper body exoskeleton and a lower limb exoskeleton. The upper body exoskeleton and the lower limb exoskeleton respectively collect the first posture data of the operator's upper body joints and the second posture data of the lower limb joints. Upper body joint mapping: The first posture data is mapped to the corresponding upper body joints of the robot to realize the upper body remote operation control of the robot; Lower limb intention recognition: Based on the second posture data, the motion intention is extracted through hybrid hierarchical logic to obtain the lower limb intention command; Lower limb joint control: Based on the lower limb intention command, a course reinforcement learning framework with random upper body joint targets as dynamic perturbations and a logarithmic decay mechanism is used as the control strategy to generate control torques for each lower limb joint of the robot, thereby realizing the lower limb teleoperation control of the robot. The control system corresponding to the control strategy at each control step The received observation state vector is ,in, The lower limb intention command includes the desired horizontal linear velocity, desired yaw rate, and desired torso height in the torso coordinate system; , These represent the base angular velocity and gravity vector of the robot in the torso coordinate system. , These represent the joint positions and velocity vectors of the entire robot. This represents the robot's lower limb motion vector from the previous step; In the current control step, the input to the control system is obtained by concatenating the observation vectors of the current step and the past 5 steps. That is, in the current control step, the control system obtains the input based on the concatenated vector. This yields the lower limb motion vector for the current step. ; Based on the lower limb motion vector Calculate the target joint position of the robot's lower limb ,for: in, This represents the natural configuration of the robot's lower limb joints. for 3D real space, The number of joints in the robot's lower limbs; After obtaining the target joint position of the lower limb Then, the control torques for each lower limb joint of the robot are generated by the PD controller, as follows: in, For the robot Control torque of each lower limb joint , The robots are respectively Proportional gain and differential gain of each lower limb joint , The robots are respectively The actual position and actual speed of each lower limb joint; The control system corresponding to the control strategy adopts an actor-critic network architecture and is trained based on a near-end policy optimization algorithm. During training, the upper body is used as a disturbance term, specifically including: Introducing a difficulty factor The amplitude boundary of random upper body perturbation is dynamically adjusted, and the difficulty factor is dynamically updated according to the average reward of the strategy to adjust the training difficulty of random upper body joint movements. Calculate the logarithmic decay parameter based on the aforementioned difficulty factor. And based on the logarithmic decay parameter Constructing the truncated exponential sampling distribution ,for: in, The initial attenuation coefficient, This is the scaling factor. This defines the boundaries of the range of motion for the upper body joints. Based on the truncation index sampling distribution Perform inverse transformation sampling to generate corresponding random upper body joint target positions. This allows for progressive training, starting with small-amplitude perturbations and progressing to kinematic limit perturbations in the later stages. Furthermore, a composite reward function, which includes speed tracking rewards and composite height rewards, is constructed to guide the control strategy to achieve lower limb stability control under upper body dynamic disturbances.

2. The whole-body teleoperation method for a humanoid robot based on joint matching and intent recognition according to claim 1, characterized in that, The extraction process of the lower limb intention command is as follows: Acquire the second attitude data, including the left thigh pitch angle. Right thigh pitch angle left lower leg pitch angle and the right lower leg tilt angle ; Feature extraction was performed on the second pose data to obtain thigh symmetry features. Symmetrical features of the lower legs Left leg flexion characteristics Characteristics of right leg flexion ; Based on the aforementioned thigh symmetry features The symmetrical features of the lower leg Using fuzzy logic, a Gaussian membership function is selected to calculate the symmetry score. ; When the symmetry score Greater than the symmetry threshold At that time, based on the left lower leg pitch angle The right lower leg pitch angle And the continuous squat depth command is obtained by using shallow, medium, and deep three-level fuzzy Gaussian membership functions. and output it as the lower limb intention command; When the symmetry score Less than or equal to the symmetry threshold When the movement is determined to be asymmetrical walking / turning, the corresponding walking / turning command is obtained by matching the walking and turning rules according to the upright / tilted state of the legs through the rule base, and then output as the lower limb intention command.

3. The method for whole-body teleoperation of a humanoid robot based on joint matching and intent recognition according to claim 2, characterized in that, The symmetry of the thighs The symmetrical features of the lower leg The left leg flexion characteristics With the aforementioned right leg flexion characteristics The extraction is specifically as follows: , , , .

4. The method for whole-body teleoperation of a humanoid robot based on joint matching and intent recognition according to claim 2, characterized in that, The process of obtaining corresponding walking / turning commands based on the leg's upright / tilted state matching walking and turning rules is as follows: Preset reference angle Adjust the pitch angle of the left thigh. The pitch angle of the right thigh The pitch angle of the left lower leg And the pitch angle of the right lower leg Respectively with reference angle Compare, and select those that are larger than the reference angle. The marked one is in an upright position, and the other marked ones are in a tilted position; When the supporting thigh and lower leg are both upright and the non-supporting thigh and lower leg are both tilted, it is determined to be a walking movement, and the obtained walking command is output as the lower limb intention command. When both the left and right thighs are upright, and one of the left and right lower legs is upright while the other is tilted, it is determined to be a turning movement, and the resulting turning command is output as the lower limb intention command.

5. The whole-body teleoperation method for a humanoid robot based on joint matching and intent recognition according to claim 1, 2, 3, or 4, characterized in that, The specific process for updating the difficulty factor is as follows: in, For average reward, This is a preset threshold.

6. The whole-body teleoperation method for a humanoid robot based on joint matching and intent recognition according to claim 1, 2, 3, or 4, characterized in that, The target position of the upper body joint Specifically: in, It is a standard uniform random variable.

7. The whole-body teleoperation method for a humanoid robot based on joint matching and intent recognition according to claim 1, 2, 3, or 4, characterized in that, The speed tracking reward is specifically as follows: in, For speed tracking rewards, For the command velocity vector, This is the actual velocity vector. Sensitivity coefficient; The composite height reward is specifically as follows: in, For compound height rewards, For the desired torso height, This refers to the actual height of the base relative to the robot's feet. For the attitude term weights, For squat range indicator, The gravity vector For compensation coefficient, The sensitivity coefficient is denoted as .

Images