Wheelchair upper limb exoskeleton robot and control method thereof

Abstract

The application provides a wheelchair upper limb exoskeleton robot and a control method thereof, the robot is composed of an electric wheelchair, a base connecting rod, an exoskeleton mechanical arm, a sensing system and a control system, the exoskeleton mechanical arm adopts a physical symmetric mirror image design, can realize quick reconstruction of left and right arms and deep coupling with the electric wheelchair, and the control method is composed of an intention recognition network, a trajectory planning algorithm, feedforward compensation and auxiliary control, can realize multi-modal signal intention recognition, bionic optimization path planning, virtual tubular constraint compliant impedance control, hand active stability compensation based on wheelchair chassis physical preview and wheelchair-exoskeleton collaborative navigation obstacle avoidance control. The wheelchair upper limb exoskeleton robot provided by the application has high structural strength, good coaxial compatibility, convenient and safe adjustment, intelligent collaborative obstacle avoidance and is suitable for auxiliary rehabilitation training and daily mobile fine life operation of patients with upper limb motor dysfunction.

Description

Technical Field

[0001] This invention relates to the field of exoskeleton robot technology, specifically to a wheelchair upper limb exoskeleton robot and its control method. Background Technology

[0002] The number of patients with upper limb motor dysfunction caused by neurological diseases such as stroke and spinal cord injury is increasing year by year. These patients often also experience lower limb mobility issues and rely on wheelchairs for daily mobility, further reducing their ability to live independently. Current research shows that high-frequency, long-term repetitive motor rehabilitation training can significantly activate brain plasticity, promote neural pathway remodeling, and restore motor function. Traditional manual rehabilitation therapy not only consumes a large amount of medical resources but also struggles to achieve precise quantitative assessment of motor parameters and customized assistance programs. Therefore, upper limb exoskeleton assistive robots have emerged as a more economical and customized rehabilitation solution, possessing enormous market value and research potential.

[0003] Existing upper limb exoskeleton robots still have some shortcomings. The following core issues urgently need to be addressed in terms of mechanical structure design, system compatibility, and collaborative operation with wheelchair mobility platforms:

[0004] Lack of bilateral reconfigurability and high cost: Clinically, paralysis in patients usually presents as unilateral hemiplegia (left or right side). Existing exoskeletons are mostly dedicated robotic arms designed for only one side (either purely left or purely right). To accommodate different patients, hospitals or families have to purchase multiple sets of equipment, significantly increasing the cost of equipment procurement and maintenance. Developing a robotic arm structure capable of physical symmetry reconfiguration within minutes and serving two purposes is imperative.

[0005] The challenge lies in adapting the arm length to the user's body type: patients of different heights have significant differences in the physical length of their upper arm and forearm. If the length of the exoskeleton linkage is fixed or cumbersome to adjust, the exoskeleton's elbow joint axis may not be precisely aligned with the physiological axis of the human elbow, resulting in pulling during flexion and extension movements. Traditional bolt adjustment methods are not only time-consuming and laborious, but also have poor adjustment accuracy and cannot perform micron-level stepless fine-tuning.

[0006] The control system is both vulnerable and rigid: conventional control strategies rely heavily on single electromyographic signals, making them susceptible to interference from fatigue and sweat; meanwhile, motion planning often uses purely geometric shortest paths, ignoring the anatomical comfort zone of the human arm. Furthermore, high-rigidity trajectory tracking control completely locks out the wearer's autonomous fine-tuning space, suppressing the patient's active motor participation.

[0007] Lack of collaborative control and environmental awareness / obstacle avoidance capabilities in mobile work scenarios: Most existing upper limb exoskeletons are fixed to static bases such as tabletops or walls, failing to meet patients' needs for free movement. Even when conventional exoskeletons are rigidly mounted on wheelchairs, the lack of a collaborative control architecture between the wheelchair chassis and the exoskeleton means that vibrations of the chassis during movement are easily transmitted to the end effector of the robotic arm, causing objects to fall or operational instability. Furthermore, in complex and dynamic indoor environments, patients find it difficult to simultaneously control the robotic arm to grasp objects and maneuver the wheelchair to avoid obstacles, greatly increasing the risk of collisions and posing serious safety hazards. Summary of the Invention

[0008] Purpose of the invention: The first purpose of this invention is to provide a wheelchair upper limb exoskeleton robot with high-precision coaxial human-machine interface and the ability to quickly reconstruct physical mirror images. The second purpose is to provide a control method for a wheelchair upper limb exoskeleton robot that integrates multimodal fusion intention recognition, cooperative navigation obstacle avoidance control, and path planning.

[0009] Technical solution: A wheelchair upper limb exoskeleton robot, comprising: an electric wheelchair, base linkages symmetrically arranged on both sides of the electric wheelchair, an exoskeleton robotic arm fixedly connected to the base linkages, a sensing system, and a control system;

[0010] The exoskeleton robotic arm includes a shoulder joint assembly, an upper arm assembly, and a forearm assembly connected in sequence. The shoulder joint assembly includes a first motor connecting plate, a first motor fixed to the first motor connecting plate, and a multi-link mechanism connected to the first motor connecting plate. The multi-link mechanism includes a first link assembly and a second link assembly, which are hinged to each other to form a bendable joint. It also includes a driven wheel coaxially connected to the hinge of the bendable joint and a second motor that drives the driven wheel to rotate. The rotation of the driven wheel is used to drive the bendable joint to bend. A third motor is provided at the end of the second link assembly. The output end of the first motor is fixedly connected to the base link. The upper arm assembly includes an upper arm link fixedly connected to the output end of the third motor and an upper arm guard fixed to the upper arm link. The forearm assembly includes a fourth motor connecting plate fixedly connected to the other end of the upper arm link, a fourth motor fixed to the fourth motor connecting plate, and a forearm link fixed to the output end of the fourth motor.

[0011] The sensing system includes a depth camera fixed above the base link, an extensor surface electromyography sensor and a flexor surface electromyography sensor fixed on the forearm link, an exoskeleton control box fixed on one side of the electric wheelchair armrest, and absolute encoders built into the first motor, second motor, third motor and fourth motor. The exoskeleton control box has an inertial measurement unit embedded in it.

[0012] The control system includes a microprocessor and a microcontroller housed in the exoskeleton control box. The microprocessor and microcontroller communicate in real time via an asynchronous transceiver. The control system is used to receive data collected by the sensor system and control the exoskeleton robotic arm and the electric wheelchair.

[0013] Preferably, the shoulder joint assembly further includes a first limiting mechanism and a second limiting mechanism. The first limiting mechanism includes a first limiting block, a first boss, and a first limiting pin. The first limiting block is fixed to one side of the first motor connecting plate, and the first boss is fixed to the other side of the first motor connecting plate and is symmetrically arranged with the first limiting block. The first limiting pin passes through the base connecting rod and is limited between the first limiting block and the first boss. The second limiting mechanism includes a second limiting block and a limiting screw. The second limiting block is fixed to one side of the third motor connecting plate. The second limiting block has symmetrically opened screw holes at both ends. The limiting screw cooperates with one of the screw holes to limit the upper arm connecting rod between the limiting screw and the second limiting block.

[0014] Preferably, the forearm assembly further includes a third limiting mechanism, which includes a limiting pin bracket fixedly connected to the forearm link, a second limiting pin disposed on the limiting pin bracket, and a second boss on the fourth motor connecting plate. After the second limiting pin is inserted into the limiting pin bracket, the forearm link is restricted to one side of the second limiting pin.

[0015] Specifically, the first linkage assembly includes a first linkage, a second linkage hinged to the other end of the first linkage, and a third linkage hinged to the first motor connecting plate and parallel to the second linkage. The second motor is fixed on the second linkage, and the output end of the second motor is fixedly connected to the drive wheel. The driven wheel is fixedly connected to the fourth linkage. The drive wheel and the driven wheel are connected by a synchronous belt. The second linkage assembly includes a fourth linkage hinged to the other end of the second linkage, a fifth linkage hinged to the other end of the third linkage and parallel to the fourth linkage, and a sixth linkage hinged to the other end of the fifth linkage. The other end of the sixth linkage is hinged to the third motor connecting plate. The third motor is fixed on the third motor connecting plate. The first linkage and the sixth linkage have the same length, the second linkage and the fifth linkage have the same length, and the third linkage and the fourth linkage have the same length. The third linkage and the fourth linkage intersect each other and are hinged at the intersection.

[0016] The present invention also provides a control method for the above-mentioned wheelchair upper limb exoskeleton robot, comprising the following steps:

[0017] S1. Use a depth camera to acquire depth images and RGB images, use extensor and flexor surface electromyography (EMG) sensors to acquire physiological EMG signals, use an inertial measurement unit to measure the speed information of the electric wheelchair in real time, and use an absolute encoder to acquire the joint angles and joint torques of the exoskeleton robotic arm to construct the original dataset.

[0018] S2. Use neural networks to extract features from the original dataset, perform feature fusion to obtain multimodal temporal features, use multimodal temporal features to identify the user's intention state, and generate a workflow state machine.

[0019] S3. Based on the current intention state of the workflow state machine, a segmented planning strategy is adopted, combined with wheelchair exoskeleton collaborative control and dynamic navigation obstacle avoidance algorithm, to generate the collaborative obstacle avoidance trajectory of the electric wheelchair and the expected movement reference trajectory of the exoskeleton hand.

[0020] S4. Real-time multi-rigid-body dynamics feedforward compensation is performed based on the kinematic model, and feedforward compensation is implemented on the exoskeleton robot arm by estimating the load weight of the exoskeleton robot arm.

[0021] S5. The tubular constraint compliant impedance control method is adopted to establish a safe channel and output control force along the desired motion reference trajectory of the exoskeleton hand. At the same time, active stabilization compensation control of the hand is implemented for the inertial disturbance generated during the movement of the electric wheelchair, and the total driving torque of the exoskeleton joint is output.

[0022] Specifically, step S2 includes: obtaining visual confidence scores, hand spatial state features, and electromyographic normalized signals from the original dataset, performing feature concatenation to obtain a multimodal temporal feature vector, inputting it into a cascaded neural network consisting of a one-dimensional convolutional neural network, a bidirectional long short-term memory network, and a self-attention mechanism module connected in sequence, outputting the probability of each intention state of the user and driving the workflow state machine to stably jump between approaching the target, grasping completion, load transfer, and releasing the hand.

[0023] Specifically, the method for obtaining the visual confidence score includes: inputting the depth image and RGB image of the current environment captured in real time by the depth camera into the object detection visual depth network model; extracting image features from the RGB image and predicting the target category confidence; calculating and classifying all interactive objects belonging to that category in the current environment; calculating the three-dimensional coordinates of each interactive object by combining the depth image and depth camera parameters; assigning a number to each interactive object; determining a unique interactive object based on the user's selection instruction based on the number; and using the three-dimensional coordinates of the interactive object as the endpoint of the optimal obstacle avoidance path. Simultaneously, the object detection visual depth network model recognizes the visual confidence scores of different gesture states in real time. Once the visual confidence score of a certain gesture state reaches the activation threshold, the object detection visual depth network model outputs a start / stop trigger signal to control the operation of the path guidance module.

[0024] Specifically, step S3 includes: In the approaching interactive object stage, firstly, the wheelchair exoskeleton collaborative control and dynamic navigation obstacle avoidance algorithm is executed. Based on the acquired depth image and the real-time motion state of the electric wheelchair, a collision-free collaborative obstacle avoidance trajectory is planned online, and the moving speed and turning angle of the electric wheelchair are dynamically adjusted. For the desired motion reference trajectory of the exoskeleton hand, an improved A* obstacle avoidance path planning algorithm is adopted to introduce a biomimetic optimization cost function. The formula for node optimization is as follows:

[0025]

[0026] In the formula: This is the cost of joint angle deviation. The cost of angular acceleration impact, For the cost of spatial path curvature, , , These are weighting coefficients;

[0027] During the operation and transportation phase, a dynamic motion primitive algorithm is used to reproduce the preset teaching trajectory using nonlinear forced terms.

[0028] Specifically, step S4 includes: in the initial static state, reading the joint static current at the horizontal calibration position of the user's arm, combining the joint torque feedback of the exoskeleton robotic arm at the same position, identifying the total mass and equivalent center of mass of the user's arm through the lever balance relationship, and constructing the upper limb gravity feedforward compensation torque, the formula of which is:

[0029]

[0030] In the formula: For the upper limb gravity feedforward compensation torque, Calculate torque for the dynamics of the exoskeleton robotic arm. The torque is calculated for the dynamics of the human arm. When the workflow state machine transitions to the point where the gripping is complete and the load is suspended, the load weight of the exoskeleton robotic arm is estimated in real time.

[0031]

[0032] In the formula: For load weight, , Let Jacobian pseudo-inverse matrix be the 4-DOF spatial base space of the exoskeleton robotic arm. This represents the net joint load torque.

[0033] The corresponding active load feedforward compensation torque is generated by using the Jacobian transpose matrix. The formula is as follows:

[0034]

[0035] In the formula: , is the spatial force vector of the load in the direction of gravity.

[0036] Specifically, active hand stabilization compensation control includes:

[0037] The acceleration and angular velocity measured by the embedded inertial measurement unit are projected onto the base coordinate system to calculate the equivalent disturbance acceleration transmitted from the wheelchair chassis disturbance to the exoskeleton hand end. :

[0038]

[0039] In the formula: For absolute translational acceleration, It is the absolute angular velocity. Angular acceleration;

[0040] Let the expected relative compensation acceleration of the hand be:

[0041]

[0042] Combined with acceleration tracking deviation:

[0043]

[0044] In the formula: This refers to the actual absolute spatial acceleration of the hand;

[0045] A hybrid feedforward and feedback architecture is used to calculate the active hand stabilization compensation joint torque. The formula is as follows:

[0046]

[0047] In the formula: The system has multiple rigid body inertia matrices. These are the proportional and integral feedback gain coefficients, respectively. This is the jerkiness feedback gain coefficient.

[0048] Beneficial effects: Compared with the prior art, the significant effects of the present invention are as follows:

[0049] 1. This invention adopts an adaptive shoulder joint design with a parallelogram closed-loop six-bar structure, and designs a quick-release mechanism and a rapid left-right symmetrical reconstruction mechanism. Compared with the bulky and fixed configuration of traditional rehabilitation robotic arms, it has a lighter wearing feel, higher human-machine axis compatibility, more convenient wheelchair non-destructive transplantation and lower maintenance costs.

[0050] 2. This invention constructs a deep neural network that integrates multimodal information such as speech, 3D vision, surface electromyography and kinematics. Compared with traditional single-sensor intention recognition methods that are susceptible to environmental interference, it effectively filters out interference such as muscle fatigue and environmental reflection, and has higher intention classification accuracy, lower response latency and smoother and more seamless grasping operation workflow.

[0051] 3. This invention combines a biomimetic cost function with dynamic motion primitives to generate a natural and smooth trajectory, and innovatively constructs a variable stiffness impedance control based on three-dimensional virtual tubular constraints. Compared with traditional rigid trajectory planning and control strategies lacking compliance protection, it provides absolute normal safety boundary protection while ensuring high biomimetic compliance, and significantly reduces the patient's muscle compensatory load.

[0052] 4. This invention proposes an active hand stabilization strategy based on chassis physical pre-aiming and state observation for wheelchair mobility scenarios. By using feedforward dynamic compensation to offset high-frequency impact disturbances in advance, it has excellent dynamic anti-interference capabilities compared with traditional upper limb exoskeletons that lack dynamic environmental adaptability. It completely eliminates the hidden dangers of hand shaking and liquid spillage during mobile operations, and greatly improves the stability of patients' fine motor skills in daily living. Attached Figure Description

[0053] Figure 1 This is a schematic diagram of the overall structure of the wheelchair upper limb exoskeleton robot according to Embodiment 1 of the present invention.

[0054] Figure 2 This is a schematic diagram of the overall structure of the left and right arms of the exoskeleton robotic arm in Embodiment 1 of the present invention.

[0055] Figure 3 This is a schematic diagram of the shoulder joint assembly in Embodiment 1 of the present invention.

[0056] Figure 4 This is a schematic diagram of the first limiting mechanism of the shoulder joint in Embodiment 1 of the present invention.

[0057] Figure 5 This is a schematic diagram of the second limiting mechanism of the shoulder joint in Embodiment 1 of the present invention.

[0058] Figure 6 This is a schematic diagram of the boom assembly in Embodiment 1 of the present invention.

[0059] Figure 7 This is a schematic diagram of the forearm assembly in Embodiment 1 of the present invention.

[0060] Figure 8 This is a flowchart of the multimodal information perception and fusion process in Embodiment 2 of the present invention.

[0061] Figure 9 This is the tubular constraint impedance control loop diagram of Embodiment 2 of the present invention.

[0062] Figure 10 This is a block diagram of hand stabilization compensation control in Embodiment 2 of the present invention. Detailed Implementation

[0063] A preferred embodiment of the present invention will be further described below with reference to the accompanying drawings.

[0064] Example 1

[0065] Please see Figures 1 to 7 As shown, this embodiment provides a wheelchair upper limb exoskeleton robot, including: an electric wheelchair 1, base connecting rods 2 symmetrically arranged on both sides of the electric wheelchair, an exoskeleton robotic arm fixedly connected to the base connecting rods 2, a sensing system, and a control system.

[0066] In this embodiment, the exoskeleton robotic arm has four degrees of freedom. It is mounted on both sides of the electric wheelchair 1 via base linkage 2 and fixed with C-clamps. The C-clamps can be quickly and flexibly clamped and fixed to the armrest linkages or frame steel pipes on both sides of the electric wheelchair, which serves as the system carrier, according to the patient's rehabilitation training or fine motor skills assistance needs for the left or right arm. This not only provides high-rigidity physical support and load transfer channels for the entire exoskeleton arm but also enables non-destructive, rapid assembly and disassembly of the robotic arm on the wheelchair chassis and seamless transfer to the opposite side. The electric wheelchair 1 chassis integrates dual-side active wheel drive motors, a chassis motion controller, wheel hub encoders, and a large-capacity power battery pack, enabling data interaction and collaborative movements with the exoskeleton robotic arm to complete intelligent composite mobile operations in a large workspace.

[0067] In this embodiment, the exoskeleton robotic arm includes a shoulder joint assembly, an upper arm assembly, and a forearm assembly connected in sequence. The main body of the shoulder joint assembly is a six-bar linkage, forming a double parallelogram closed-loop kinematic chain, specifically including a first motor connecting plate 3, a first motor 4 fixed on the first motor connecting plate 3, a first link 5 hinged to the first motor connecting plate 3, a second link 6 hinged to the other end of the first link 5, a third link 7 hinged to the first motor connecting plate 3 and parallel to the second link 6, and a second link 7 fixed on the second link 6. Motor 8, drive wheel 9 fixedly connected to the output end of the second motor 8, fourth link 10 hinged to the other end of the second link 6, driven wheel 11 fixedly connected to the fourth link 10, fifth link 12 hinged to the other end of the third link 7 and parallel to the fourth link 10, sixth link 13 hinged to the other end of the fifth link 12, third motor connecting plate 14 hinged to the other end of the sixth link 13, third motor 15 fixed on the third motor connecting plate 14, and the output end of the first motor 4 fixedly connected to the base link 2.

[0068] In this embodiment, the first link 5 and the sixth link 13 have the same length, the second link 6 and the fifth link 12 have the same length, and the third link 7 and the fourth link 10 have the same length. The third link 7 and the fourth link 10 are hinged at their intersection. This design aims to achieve a symmetrical, parallel, multi-stage closed-loop transmission. The upper part, consisting of the first link 5, the second link 6, and the third link 7, forms a first parallelogram closed-loop transmission chain. The lower part, consisting of the fourth link 10, the fifth link 12, and the sixth link 13, forms a second parallelogram closed-loop transmission chain. The driving wheel 9 and the driven wheel 11 are connected by a synchronous belt 16, directly driving the lower second parallelogram closed-loop transmission chain to swing. Through the symmetrical interference and geometric constraints of the closed-loop mechanism, the three shoulder rotation axes of the exoskeleton robotic arm are always collinear and intersect at a single point in space, ensuring that the virtual motion center highly coincides with the physiological rotation center of the human glenohumeral joint, effectively avoiding the generation of parasitic interactive forces due to motion misalignment.

[0069] Please refer to Figure 4 and Figure 5 The shoulder joint assembly also includes a first limiting mechanism and a second limiting mechanism. The first limiting mechanism includes a first limiting block 31, a first boss 32, and a first limiting pin 33. The first limiting block 31 is fixed to one side of the first motor connecting plate 3, and the first boss 32 is fixed to the other side of the first motor connecting plate 3 and is symmetrically arranged with the first limiting block 31. The first limiting pin 33 passes through the base connecting rod 2 and is limited between the first limiting block 31 and the first boss 32. The first motor connecting plate 3 and the base connecting rod 2 are fixedly connected by the first shoulder joint protective shell 41. The second limiting mechanism includes a second limiting block 141 and a limiting screw 142. The second limiting block 141 is fixed to one side of the third motor connecting plate 14. The second limiting block 141 has two screw holes symmetrically opened at both ends. The limiting screw 142 cooperates with one of the screw holes to limit the upper arm connecting rod 17 between the limiting screw 142 and the second limiting block 141.

[0070] Please refer to Figure 5 and Figure 6 In this embodiment, the boom assembly includes a boom link 17 fixedly connected to the output end of the third motor 15 via a boom connecting plate 174, and a boom guard 18 fixed to the boom link 17. The boom guard 18 is fixed by a boom guard fixing block 176 and bolts. In this embodiment, a steplessly adjustable boom link 17 is used, including an outer link 171, a locking device 172, and an inner link 173. One end of the inner link 173 is nested inside the outer link 171. Rotating the locking device 173 generates an inward radial clamping force, thereby locking the relative positions of the outer link 171 and the inner link 173. To further prevent relative circumferential sliding between the inner and outer links, a limiting groove is provided on the inner link 173, and multiple limiting holes are provided on the outer link 171. Circumferential locking can be achieved by cooperating with the third limiting pin 175.

[0071] Please refer to Figure 7 In this embodiment, the forearm assembly includes a fourth motor connecting plate 19 fixedly connected to the other end of the upper arm connecting rod 17, a fourth motor 20 fixed on the fourth motor connecting plate 19, and a forearm connecting rod 21 fixed on the output end of the fourth motor 20. The forearm connecting rod 21 is fixedly connected to the forearm brace 26. A first forearm joint protective shell 201 is provided on the outside of the fourth motor 20, and a second forearm joint protective shell 202 is provided on the outside of the forearm connecting rod 21. The first forearm joint protective shell 201 and the second forearm joint protective shell 202 are fixedly connected to each other. The forearm assembly also includes a third limiting mechanism. The third limiting mechanism includes a limiting pin bracket 211 fixedly connected to the forearm connecting rod 21, a second limiting pin 212 provided on the limiting pin bracket 211, and a second boss 191 on the fourth motor connecting plate 19. After the second limiting pin 212 is inserted into the limiting pin bracket 211, the range of motion of the forearm connecting rod 21 is limited to one side of the second limiting pin 212.

[0072] Please refer to Figure 1 and Figure 7 As shown, in this embodiment, the sensing system includes a depth camera 22 fixed above the base link 2, an extensor side surface electromyography sensor 23 and a flexor side surface electromyography sensor 24 fixed on the forearm link 21, an exoskeleton control box 25 fixed on one side of the electric wheelchair armrest, and absolute encoders built into the first motor 4, the second motor 8, the third motor 15 and the fourth motor 20. The exoskeleton control box 25 has an inertial measurement unit embedded in it.

[0073] The sensing system provided in this embodiment deeply couples the user's physiological intentions, the state of the robotic arm, and the environmental spatial perception of the electric wheelchair chassis, constructing a complete "human-machine-environment" fused information network. A depth camera 22 is mounted above the base linkage 2 with a fixed downward tilt at a preset angle, using high frequency to capture the depth point cloud of the external working environment and target objects; extensor-side surface electromyography (EMG) sensors 23 and flexor-side surface EMG sensors 24 are used to extract the normalized intensity signals of two-channel surface EMG signals from the forearm extensor and flexor muscles in real time; an exoskeleton control box 25, rigidly locked to the armrest linkage of the electric wheelchair 1, integrates a high-performance microprocessor, an embedded microcontroller development board, and a high-sensitivity inertial measurement unit (IMU, used for active anti-shake stabilization control of chassis driving disturbances and high-frequency vibrations); the first motor 4, the second motor 8, the third motor 15, and the fourth motor 20 all have built-in integrated absolute encoders for high-frequency acquisition of joint rotation angles. In this embodiment, the motors at each joint do not contain additional torque sensors. Instead, the real-time feedback torque of each active joint is indirectly calculated and obtained by acquiring the drive current feedback of each joint motor and multiplying it by its corresponding motor torque constant.

[0074] The control system in this embodiment mainly consists of a high-performance microprocessor and an embedded microcontroller cascaded together. Both exchange data bidirectionally with each motor driver in real time via a high-speed asynchronous transceiver (UART) at an update frequency of 1 kHz. The high-performance microprocessor is primarily used for high-throughput multi-source intent prediction, advanced path planning for wheelchair-exoskeleton collaboration, and environmental obstacle avoidance and navigation calculations. The embedded microcontroller is used for low-level high-frequency kinematics and dynamics calculations, gravity / load feedforward compensation, and, based on the initial absolute joint angles of each active joint read during system power-on self-test, automatically determines whether the exoskeleton arm is currently mounted on the left or right side of the wheelchair using an initial angle determination algorithm. It adaptively reconstructs and loads the kinematic parameters of the corresponding side from memory, coupled with the physical dimensions of the wheelchair chassis, thus completely eliminating the risk of instability introduced by human misconfiguration from a hardware topology perspective.

[0075] The wheelchair-mounted upper limb exoskeleton robot provided in this embodiment has excellent flexibility and reusability. It can complete the physical symmetrical mirror reconstruction of the left / right exoskeleton robotic arms within 3 minutes. The reconstruction and debugging steps are as follows:

[0076] Step 1: The rehabilitation physician loosens the eccentric quick-release wrench of the C-clamp, causing the C-clamp to detach from the armrest link on one side of the electric wheelchair 1. The physician then moves the entire exoskeleton robotic arm to the armrest tube link on the opposite side and re-tightens the C-clamp.

[0077] Step 2: Then pull the first limiting pin 33 on the base connecting rod 2 outward to retract and release the limit. Rotate the base connecting rod 2 relative to the first limiting block 31 installed on the first motor connecting plate 3 to a symmetrical mirror angle. Release the first limiting pin 33 so that it pops out under the spring return force and inserts into the symmetrical hole to lock the new limit. It also cooperates with the first boss 32 to physically lock the new safety limit.

[0078] Step 3: Unscrew the limit screw 142 and re-screw it into the other reserved screw hole on the symmetrical mirror side of the second limit block 141 and tighten it to mechanically lock the new safe lifting movement limit of the boom.

[0079] Step 4: Pull the second limit pin 212 outward to retract it and release the limit. After rotating the forearm connecting rod 21, release the second limit pin 212 at the corresponding limit hole on the other side so that it pops out automatically, readjust the elbow flexion and extension stroke and lock it within the new safe range.

[0080] After reconstruction, when the system is initially powered on, the embedded microcontroller reads the initial absolute angle of the joint from the absolute encoder integrated inside the motor at each joint, automatically determines the current assembly configuration (left or right) of the exoskeleton robotic arm mounted on the electric wheelchair, and automatically loads the corresponding kinematic DH parameter table and dynamic calibration matrix model parameters, achieving rapid and zero-threshold reuse and completely eliminating the risk of instability caused by conflicts in manually written program configurations.

[0081] Example 2

[0082] This embodiment provides a control method for the wheelchair upper limb exoskeleton robot provided in Embodiment 1, including the following steps:

[0083] Step 1: Use embedded microcontrollers and high-performance microprocessors to collect and read feedback status data from various sensors and motors in real time.

[0084] Specifically, this embodiment acquires depth images of the target working environment from a depth camera, RGB images and cloud map data, and physiological electromyographic signals (normalized signals from the extensor and flexor sides of the forearm) from surface electromyography sensors. Normalized signal with flexor side The exoskeleton control box contains an embedded inertial measurement unit (IMU) that collects the original linear acceleration and angular velocity of the electric wheelchair chassis, the joint angle vectors and joint rotational speeds from the absolute encoders built into each motor, and the real-time feedback torque data of each joint obtained by mapping the drive current and torque constant of each joint motor.

[0085] Step 2: Utilize a high-performance microprocessor to run a cascaded deep network that integrates physiological and physical signals at high frequency, identify the user's grasping intent state, and drive the underlying workflow state machine to jump stably.

[0086] Please refer to Figure 8As shown, this embodiment employs a joint redundant input network configured with a voice module, depth camera, surface electromyography sensor, and robotic arm kinematics state. After the system is powered on, the voice module first captures the user's category selection command and transmits it to the target selection module, establishing the desired object category (this is just a typical example; manual typing, eye-tracking, and other input methods are also feasible). The depth camera acquires environmental depth and RGB images in real time, inputting them into a YOLOv11 object detection visual depth network model running on a high-performance microprocessor. The YOLOv11 model extracts image features from the RGB images and performs target category confidence prediction, calculating, extracting, and classifying all specific interactive objects in the environment belonging to that category. This, combined with the depth... The image and camera intrinsic parameters calculate the 3D coordinates of each object and assign a number to each object. Then, the voice module captures the user's number selection command again and transmits it to the target selection module to establish and lock onto a unique target object. The 3D coordinates of this target object are then sent to the path planning module as the endpoint for offline optimal obstacle avoidance path calculation. Simultaneously, the YOLOv11 model identifies the visual confidence scores of the hand in real-time when it is approaching, clenched, or open. Once the visual confidence score of a specific gesture reaches the activation threshold, the YOLOv11 model outputs a start / stop trigger signal to control the operation of the path guidance module. After path guidance is initiated, the robotic arm will guide the user's hand along the desired trajectory. During this process, the electromyographic intensity characteristics of the forearm extensor and flexor sides will be collected. , The system calculates the real-time 3D spatial coordinates and velocity vectors of the hand (a total of 6 features) from the forward kinematics of the robotic arm. These features are then synchronized with the visual confidence score output by the YOLOv11 model over a high-frequency time scale using a timestamp synchronization mechanism to construct a unified-dimensional multimodal temporal feature vector. This multimodal temporal feature vector is then fed into a grasping intent model (i.e., a cascaded deep network) for cascaded feature fusion and state classification. The cascaded deep network sequentially includes: a one-dimensional convolutional neural network (containing two one-dimensional convolutional layers, a normalization layer, and a ReLU activation function) for extracting cross-channel coupling features; a bidirectional long short-term memory network (Bi-LSTM) for capturing temporal evolution trends; and a self-attention mechanism module for adaptively weighting keyframes of feature mutations. Finally, a fully connected layer and a Softmax classifier output the probability values ​​of the current grasping state, driving the workflow state machine to stably transition between the approximation phase, grasping completion, load transfer, and hand release with high accuracy and low latency. When the grip intention model calculates that the wearer has completed a firm grip on the object, the grip state output starts the control flow, automatically triggering the load compensation module to load the joint load feedforward torque to assist in lifting the heavy object; if the release command is detected, the state machine is reset and outputs a stop signal to unload the load compensation torque, realizing adaptive flow driven by multi-source information perception throughout the process.

[0087] Step 3: Based on the current intention state of the workflow state machine, adopt a segmented planning strategy and execute macro-micro dual-layer motion collaborative control and autonomous navigation obstacle avoidance of the wheelchair mobile chassis and exoskeleton robotic arm.

[0088] In the approach phase to the heavy object / target, in order to achieve smooth operation in a large workspace and eliminate blind spots, the bionic obstacle avoidance path of the exoskeleton robotic arm and the collaborative local navigation path of the intelligent electric wheelchair chassis are calculated simultaneously.

[0089] On the one hand, exoskeleton robotic arm path planning introduces a spatial optimization cost function that includes joint comfort deviations, angular acceleration impacts, and spatial path curvature. The improved A* obstacle avoidance path planning algorithm is used for 3D mesh node optimization, and its total cost optimization formula is:

[0090]

[0091] In the formula: The actual geometric path mesh cost, This is for heuristic cost estimation.

[0092] The improved A* obstacle avoidance path planning algorithm introduces a biomimetic optimization cost function that incorporates joint comfort bias, angular acceleration impact, and spatial path curvature. Perform 3D mesh node optimization:

[0093]

[0094] In the formula: This is the cost of joint angle deviation based on the statistically accepted angle range of human comfort (i.e., the normalized maximum value of the real-time angle of each joint relative to the midpoint of the comfort range). The joint angular acceleration impact cost is based on the maximum ratio of the differential angular acceleration between adjacent search nodes to the physiological safety limit. The normalized spatial path curvature cost is used to reflect the degree of geometric deflection and tortuosity of the spatial path. These are the weighting coefficients for the corresponding optimization terms.

[0095] On the other hand, the system performs real-time wheelchair-exoskeleton collaborative motion guidance and chassis local obstacle avoidance navigation control: the collaborative target tracking and navigation control method running on a high-performance microprocessor first uses a depth camera installed above the exoskeleton base linkage to obtain the three-dimensional spatial relative position vector between the wheelchair chassis and the target object in real time, and calculates the deviation of the object relative to the workspace; at the same time, the central controller integrates the environmental depth local cloud map established by the depth camera, and adopts a local obstacle avoidance path planning algorithm based on artificial potential field, setting the target object as the center of the gravitational field that generates attraction, and setting the obstacles identified by the depth camera as the center of the repulsive field that generates repulsion.

[0096] Let the current two-dimensional spatial position of the chassis be... The target location is The gravity acting on the chassis The calculation formula is:

[0097]

[0098] In the formula, Let be the gravitational gain coefficient. Assume the obstacle's position is... , The distance between the chassis and the obstacle. For the radius of influence of the obstacle, when At that time, the repulsive force on the chassis The calculation formula is:

[0099]

[0100] In the formula, The repulsive force gain coefficient, Let be the gradient vector of the distance.

[0101] The resultant force vector is obtained by superimposing the gravitational and repulsive forces: .

[0102] It automatically plans a smooth, collision-free approach and local avoidance trajectory for the intelligent electric wheelchair chassis, and then uses the resultant force vector... The direction and size are calculated online in real time to determine the expected linear velocity and yaw rate of the electric wheelchair. This drives the electric wheelchair chassis to actively and smoothly adjust the lateral / longitudinal displacement and the yaw angle of the vehicle body, so as to autonomously avoid obstacles and approach the target. This ensures that the target object is always stably kept within the optimal detection field of view of the depth camera and the optimal high operability operation envelope and comfort range of the exoskeleton robotic arm, realizing the macro-micro space coordination of "using the vehicle instead of the arm".

[0103] If the vision system suddenly detects a close-range static dangerous barrier or a sudden dynamic emergency obstacle, or other high-risk emergency collision scenario, the chassis perception layer automatically triggers the active safety mechanism and forcibly issues a safety braking command to the wheelchair parking system, so that the wheelchair can quickly park and stop, thus effectively ensuring the absolute safety of the human and machine.

[0104] During the operation and transportation phase, the dynamic motion primitive algorithm uses nonlinear forcing terms to learn and accurately reproduce the fine living skills (such as drinking, delivering medicine or transporting materials) taught by the rehabilitation therapist. Furthermore, when the depth camera captures the latest three-dimensional target position of the patient's mouth or the target operation point, the algorithm adjusts the target position vector online to achieve smooth generalization, adaptive stepless scaling, and stable movement of the hand's expected relative motion control reference trajectory.

[0105] Step 4: To help patients with weakness or fatigue overcome the gravitational torque of their own arms and the heavy load when holding objects, the embedded microcontroller performs real-time high-frequency multi-rigid-body dynamics feedforward compensation in the wearable state.

[0106] When the system is initially stationary, the command motor driver moves the user's arm to a horizontally calibrated position and reads the joint resting current. Combined with the joint torque feedback of the exoskeleton robotic arm in the same position, the total mass and equivalent center of mass of the human arm are identified through lever balance relationship, thereby constructing the upper limb gravity feedforward compensation torque. The expression is:

[0107]

[0108] In the formula: and The torque values ​​represent the dynamic calculation torques of the exoskeleton and the human arm, respectively.

[0109] When the state machine transitions to the state of complete gripping and the load is suspended, the system does not require cumbersome and easily damaged external tension and compression sensors. It directly estimates the actual load weight of the load in real time using the following formula. The formula is:

[0110]

[0111] In the formula: , Let Jacobian pseudo-inverse matrix be the base space of the 4-DOF space of the robotic arm. The values ​​were calculated from the drive current of each active joint motor, after removing the static friction of the motor itself and the aforementioned self-gravity torque. The net joint load torque after that.

[0112] Finally, the system generates the corresponding active load feedforward compensation torque through the Jacobian transpose matrix. :

[0113]

[0114] In the formula: , where is the spatial force vector of the load in the direction of gravity. Through this active feedforward gravity and load feedforward compensation, the activation level of the wearer's upper arm electromyographic signal is significantly reduced, effectively improving the transparency of motion assistance in human-computer interaction.

[0115] Step 5: Using a tubular constraint compliant impedance control method, a three-dimensional virtual safety tubular channel is established along the desired motion reference trajectory. Low stiffness compliant free operation is provided inside the channel, and an exponentially high impedance absolute safety limit constraint is provided at the channel boundary. At the same time, on-demand auxiliary or damping control is implemented in the tangential direction.

[0116] Please refer to Figure 9 First, along the desired three-dimensional reference path Establish its boundary radius as Virtual 3D security channel:

[0117]

[0118] In the formula: The path projection parameters are the closest to the actual position of the current hand end.

[0119] Establish the desired tangential unit vector The actual tangential velocity of the hand is The normal deviation vector is .

[0120] Spatial guiding force on the hand Decoupling is achieved through tangential forces (including auxiliary / damping forces). With normal constraint force It is formed by the superposition of vectors. Among them, the tangential guiding force scalar In assisted mode, the speed is adjusted by an assist-as-needed loop based on the deviation between the walking speed and the set upper and lower speed threshold range; in resistance rehabilitation mode, resistance muscle strength training is implemented through tangential damping resistance.

[0121] In the normal direction, the control force consists of the tangential force and the normal constraint force. It is formed by the superposition of vectors, in which the normal constraint force An exponential variable stiffness impedance control strategy is adopted:

[0122]

[0123] In the formula: The basic normal stiffness coefficient, The normal damping coefficient is... This is the dimensionless normal deviation rate.

[0124] Stiffness amplification factor The calculation formula is:

[0125]

[0126] In the formula: The minimum stiffness factor of the boundary. The standard stiffness factor of the boundary. The curvature constant is used to control the steepness of the stiffness curve. To limit the output amplitude saturation function, the stiffness factor is ensured to be limited within a preset safe and stable range. The above stiffness parameter settings ensure that the stiffness is extremely low in the core area of ​​the channel axis, allowing the wearer to freely deviate and make fine adjustments, while the stiffness increases sharply instantaneously when crossing the boundary, forcibly bouncing the hand back.

[0127] Furthermore, during movement outdoors and indoors with the electric wheelchair chassis, the exoskeleton system is highly susceptible to high-frequency physical vibrations and impacts from uneven road surfaces, speed bumps, or other bumps. Especially in scenarios requiring fine motor skills such as holding easily spilled objects (e.g., water or medicine cups) while simultaneously controlling the wheelchair, the high-frequency impact disturbances from the chassis are transmitted to the exoskeleton's robotic arm, easily causing spills or severe hand tremors. To address this issue, active hand stabilization compensation control is employed; please refer to [reference needed]. Figure 10 The calculation process is as follows:

[0128] The exoskeleton control box is rigidly fixed to the armrest linkage of the electric wheelchair. The embedded high-frequency inertial measurement unit (IMU) provides advance physical pre-aiming of external translational and rotational impacts. The measured acceleration and angular velocity are projected and transformed to the base coordinate system and gravity is removed to obtain the absolute translational acceleration. Absolute angular velocity With angular acceleration .

[0129] Using a multi-rigid-body spatial kinematics transfer model, the equivalent absolute perturbation acceleration transmitted to the hand end is calculated. for:

[0130]

[0131] In the formula: The current three-dimensional spatial coordinate position of the hand end relative to the exoskeleton base coordinate system is obtained in real time by forward kinematics.

[0132] Desirable relative compensation acceleration of the hand A hybrid feedforward and feedback architecture is used to calculate active stabilization compensation joint torque. :

[0133]

[0134] In the formula: The system has multiple rigid body inertia matrices. Let Jacobian pseudo-inverse matrix be the robotic arm. For Jacobian matrices, These are the proportional and integral feedback gain coefficients, respectively. This represents the jerk feedback gain coefficient. To track acceleration deviation, This represents the actual absolute spatial acceleration of the hand as observed by the joint absolute encoder.

[0135] The feedforward path achieves proactive dynamic compensation for the equivalent virtual inertial mass of the hand operation space, while the feedback path introduces jerk feedback phase advance compensation feedback output. This is used to offset communication and computing delays. This ensures that, under combined conditions of severe chassis vibration and variable stiffness impedance self-management, the exoskeleton arm's end effector always maintains precise trajectory safety boundary constraints and excellent spatial anti-shake and anti-bumping capabilities.

Claims

1. A wheelchair-mounted upper limb exoskeleton robot, characterized in that, include: Electric wheelchair (1), base connecting rods (2) symmetrically arranged on both sides of the electric wheelchair (1), exoskeleton robotic arm, sensing system and control system fixedly connected to the base connecting rods (2); The exoskeleton robotic arm includes a shoulder joint assembly, an upper arm assembly, and a forearm assembly connected in sequence. The shoulder joint assembly includes a first motor connecting plate (3), a first motor (4) fixed on the first motor connecting plate (3), and a multi-link mechanism connected to the first motor connecting plate (3). The multi-link mechanism includes a first link assembly and a second link assembly, which are hinged to each other to form a bendable joint. It also includes a driven wheel (11) coaxially connected to the hinge of the bendable joint and a second motor (8) that drives the driven wheel (11) to rotate. The rotation of the driven wheel (11) drives the bendable joint to rotate. The bending joint is bent, and the end of the second link assembly is provided with a third motor (15); the output end of the first motor (4) is fixedly connected to the base link (2); the upper arm assembly includes an upper arm link (17) fixedly connected to the output end of the third motor (15) and an upper arm guard (18) fixed on the upper arm link (17); the lower arm assembly includes a fourth motor connecting plate (19) fixedly connected to the other end of the upper arm link (17), a fourth motor (20) fixed on the fourth motor connecting plate (19), and a lower arm link (21) fixed on the output end of the fourth motor (20). The sensing system includes a depth camera (22) fixed above the base link (2), an extensor side surface electromyography sensor (23) and a flexor side surface electromyography sensor (24) fixed on the forearm link (21), an exoskeleton control box (25) fixed on one side of the armrest of the electric wheelchair (1), and absolute encoders built into the first motor (4), the second motor (8), the third motor (15), and the fourth motor (20). The exoskeleton control box (25) is embedded with an inertial measurement unit. The control system includes a microprocessor and a microcontroller installed in the exoskeleton control box (25). The microprocessor and the microcontroller communicate in real time via an asynchronous transceiver. The control system is used to receive data collected by the sensor system and control the exoskeleton robotic arm and the electric wheelchair (1).

2. The wheelchair-mounted upper limb exoskeleton robot according to claim 1, characterized in that: The shoulder joint assembly also includes a first limiting mechanism and a second limiting mechanism. The first limiting mechanism includes a first limiting block (31), a first boss (32), and a first limiting pin (33). The first limiting block (31) is fixed on one side of the first motor connecting plate (3), and the first boss (32) is fixed on the other side of the first motor connecting plate (3) and is symmetrically arranged with the first limiting block (31). The first limiting pin (33) passes through the base connecting rod (2) and is limited between the first limiting block (31) and the first boss (32). The second limiting mechanism includes a second limiting block (141) and a limiting screw (142). The second limiting block (141) is fixed on one side of the third motor connecting plate (14). The second limiting block (141) has symmetrical screw holes at both ends. The limiting screw (142) cooperates with one of the screw holes to limit the upper arm connecting rod (17) between the limiting screw (142) and the second limiting block (141).

3. The wheelchair-mounted upper limb exoskeleton robot according to claim 1, characterized in that: The forearm assembly also includes a third limiting mechanism, which includes a limiting pin bracket (211) fixedly connected to the forearm link (21), a second limiting pin (212) disposed on the limiting pin bracket (211), and a second boss (191) on the fourth motor connecting plate (19). After the second limiting pin (212) is inserted into the limiting pin bracket (211), the forearm link (21) is restricted to one side of the second limiting pin (212).

4. The wheelchair-mounted upper limb exoskeleton robot according to claim 1, characterized in that: The first linkage assembly includes a first link (5), a second link (6) hinged to the other end of the first link (5), and a third link (7) hinged to the first motor connecting plate (3) and parallel to the second link (6). The second motor (8) is fixed on the second link (6), and the output end of the second motor (8) is fixedly connected to the driving wheel (9). The driven wheel (11) is fixedly connected to the fourth link (10). The driving wheel (9) and the driven wheel (11) are connected by a synchronous belt (16). The second linkage assembly includes a fourth link (10) hinged to the other end of the second link (6), and a third link (7) hinged to the first motor connecting plate (3) and parallel to the second link (6). 7) The other end is hinged to the fifth link (12) and parallel to the fourth link (6), and the other end is hinged to the sixth link (13). The other end of the sixth link (13) is hinged to the third motor connecting plate (14). The third motor (15) is fixed on the third motor connecting plate (14). The first link (5) and the sixth link (13) are the same length. The second link (6) and the fifth link (12) are the same length. The third link (7) and the fourth link (10) are the same length. The third link (7) and the fourth link (8) intersect each other and are hinged at the intersection.

5. A control method for a wheelchair-mounted upper limb exoskeleton robot according to any one of claims 1-4, characterized in that, Includes the following steps: S1. Use a depth camera to acquire depth images and RGB images, use extensor and flexor surface electromyography (EMG) sensors to acquire physiological EMG signals, use an inertial measurement unit to measure the speed information of the electric wheelchair in real time, and use an absolute encoder to acquire the joint angles and joint torques of the exoskeleton robotic arm to construct the original dataset. S2. Use neural networks to extract features from the original dataset, perform feature fusion to obtain multimodal temporal features, use multimodal temporal features to identify the user's intention state, and generate a workflow state machine. S3. Based on the current intention state of the workflow state machine, a segmented planning strategy is adopted, combined with wheelchair exoskeleton collaborative control and dynamic navigation obstacle avoidance algorithm, to generate the collaborative obstacle avoidance trajectory of the electric wheelchair and the expected movement reference trajectory of the exoskeleton hand. S4. Real-time multi-rigid-body dynamics feedforward compensation is performed based on the kinematic model, and feedforward compensation is implemented on the exoskeleton robot arm by estimating the load weight of the exoskeleton robot arm. S5. The tubular constraint compliant impedance control method is adopted to establish a safe channel and output control force along the desired motion reference trajectory of the exoskeleton hand. At the same time, active stabilization compensation control of the hand is implemented for the inertial disturbance generated during the movement of the electric wheelchair, and the total driving torque of the exoskeleton joint is output.

6. The control method for the wheelchair upper limb exoskeleton robot according to claim 5, characterized in that, Step S2 includes: obtaining visual confidence scores, hand spatial state features, and electromyographic normalized signals from the original dataset; performing feature concatenation to obtain a multimodal temporal feature vector; inputting it into a cascaded neural network consisting of a one-dimensional convolutional neural network, a bidirectional long short-term memory network, and a self-attention mechanism module connected in sequence; outputting the probability of each intention state of the user; and driving the workflow state machine to stably jump between approaching the target, grasping completion, load transfer, and releasing the hand.

7. The control method for the wheelchair upper limb exoskeleton robot according to claim 6, characterized in that, The method for obtaining the visual confidence score includes: inputting the depth image and RGB image of the current environment acquired in real time by the depth camera into the target detection visual depth network model; extracting image features from the RGB image and predicting the target category confidence; calculating and classifying all interactive objects belonging to that category in the current environment; calculating the three-dimensional coordinates of each interactive object by combining the depth image and depth camera parameters; assigning a number to each interactive object; determining a unique interactive object based on the user's number selection instruction; and using the three-dimensional coordinates of the interactive object as the endpoint of the optimal obstacle avoidance path. Simultaneously, the target detection visual depth network model recognizes the visual confidence scores of different gesture states in real time. Once the visual confidence score of a certain gesture state reaches the activation threshold, the target detection visual depth network model outputs a start / stop trigger signal to control the operation status of the path guidance module.

8. The control method for the wheelchair upper limb exoskeleton robot according to claim 7, characterized in that, Step S3 includes: In the approaching interactive object stage, firstly, the wheelchair exoskeleton collaborative control and dynamic navigation obstacle avoidance algorithm is executed. Based on the acquired depth image and the real-time motion state of the electric wheelchair, a collision-free collaborative obstacle avoidance trajectory is planned online, and the moving speed and turning angle of the electric wheelchair are dynamically adjusted. For the desired motion reference trajectory of the exoskeleton hand, an improved A* obstacle avoidance path planning algorithm is adopted to introduce a biomimetic optimization cost function. The formula for node optimization is as follows: In the formula: This is the cost of joint angle deviation. The cost of angular acceleration impact, For the cost of spatial path curvature, , , These are weighting coefficients; During the operation and transportation phase, a dynamic motion primitive algorithm is used to reproduce the preset teaching trajectory using nonlinear forced terms.

9. The control method for the wheelchair upper limb exoskeleton robot according to claim 8, characterized in that, Step S4 includes: In the initial static state, reading the joint static current at the horizontal calibration position of the user's arm; combining this with the joint torque feedback of the exoskeleton robotic arm at the same position; identifying the total mass and equivalent center of mass of the user's arm through lever balance; and constructing the upper limb gravity feedforward compensation torque, as shown in the formula: In the formula: For the upper limb gravity feedforward compensation torque, Calculate torque for the dynamics of the exoskeleton robotic arm. The torque is calculated for the dynamics of the human arm. When the workflow state machine transitions to the point where the gripping is complete and the load is suspended, the load weight of the exoskeleton robotic arm is estimated in real time. In the formula: For load weight, , Let Jacobian pseudo-inverse matrix be the 4-DOF spatial base space of the exoskeleton robotic arm. This represents the net joint load torque. The corresponding active load feedforward compensation torque is generated by using the Jacobian transpose matrix. The formula is as follows: In the formula: , is the spatial force vector of the load in the direction of gravity.

10. The control method for the wheelchair upper limb exoskeleton robot according to claim 9, characterized in that, The active hand stabilization compensation control includes: The acceleration and angular velocity measured by the embedded inertial measurement unit are projected onto the base coordinate system to calculate the equivalent disturbance acceleration transmitted from the wheelchair chassis disturbance to the exoskeleton hand end. : In the formula: For absolute translational acceleration, It is the absolute angular velocity. Angular acceleration; Let the expected relative compensation acceleration of the hand be: Combined with acceleration tracking deviation: In the formula: This refers to the actual absolute spatial acceleration of the hand; A hybrid feedforward and feedback architecture is used to calculate the active hand stabilization compensation joint torque. The formula is as follows: In the formula: The system has multiple rigid body inertia matrices. These are the proportional and integral feedback gain coefficients, respectively. This is the jerkiness feedback gain coefficient.

Images