Self-adaptive closed-loop exoskeleton system based on multi-modal brain-computer interface

Abstract

The invention relates to a self-adaptive closed-loop exoskeleton system based on a multi-modal brain-computer interface. The self-adaptive closed-loop exoskeleton system comprises a) a multi-modal signal acquisition unit; b) an adaptive neural decoding algorithm module; c) the three-joint exoskeleton follows the motion intention in real time through a hierarchical control strategy to control exoskeleton joints to move, and gait parameters are dynamically adjusted; d) a redundancy switching unit which is switched to an EEG full-function mode within 200ms when the ECoG signal is abnormal; and e) a spinal cord electrical stimulation module which is used for providing closed-loop auxiliary stimulation when the electroencephalogram signal is weak. High-precision and high-reliability motion control can be provided, the self-adaptive complex environment interaction capability is achieved, nerve function remodeling and rehabilitation can be promoted, and a safe solution is provided for motion control of a spinal cord injury patient.

Description

Technical Field

[0001] This invention relates to the field of brain-computer interface technology, and in particular to an adaptive closed-loop exoskeleton system based on a multimodal brain-computer interface. Background Technology

[0002] Spinal cord injury disrupts communication between the brain and the spinal cord's motor centers, leading to lower limb paralysis (such as paraplegia or quadriplegia). Current technologies rely on invasive surgery (such as electrical epidural stimulation, EES) or passive rehabilitation training, which have the following drawbacks: Lack of natural control: EES requires pre-programmed stimulation and cannot respond to the patient's motor intentions in real time; Poor adaptability: It cannot adapt to complex terrain (such as slopes, obstacles) or adjust gait parameters (such as stride length and speed) in real time; Long-term dependence: Patients need to wear external devices continuously, and neurological function recovery is limited (only restoring assisted walking ability).

[0003] Brain-computer interface (BCI) technology faces technical bottlenecks in the treatment of spinal cord injuries. Non-invasive BCI (such as EEG) has high signal noise and low decoding accuracy, typically <80%. Invasive BCI (such as ECoG) requires craniotomy to implant electrodes, resulting in insufficient long-term stability, with signal attenuation typically >0.03 dB / day.

[0004] It can be seen that existing spinal cord injury rehabilitation exoskeletons have three major defects: insufficient EEG signal decoding accuracy, traditional EEG is easily interfered with; single degree of freedom control cannot meet the needs of complex movements; and passive training mode inhibits neuroplasticity. Summary of the Invention

[0005] The purpose of this invention is to address the problems of unnatural motor control, poor adaptability, and insufficient neurological function recovery in spinal cord injury patients in the prior art. It provides a brain-controlled neurorehabilitation device based on an EEG signal-adaptive algorithm-exoskeleton closed-loop system to achieve: natural motor control: real-time decoding of the patient's motor intentions (such as stepping or turning) to directly drive the exoskeleton; adaptive environmental interaction: dynamic adjustment of gait parameters (step frequency, step length) to cope with complex terrain; and neurological function remodeling: promoting spinal cord neuroplasticity through closed-loop training, gradually reducing reliance on exoskeleton assistance.

[0006] To achieve the above objectives, the present invention provides an adaptive closed-loop exoskeleton system based on a multimodal brain-computer interface, comprising:

[0007] a) A multimodal signal acquisition unit, comprising an invasive ECoG stereotactic deep brain electrode implanted in the lower limb functional area of ​​the motor cortex and a non-invasive dry electrode EEG cap. The invasive ECoG stereotactic deep brain electrode is a cylindrical electrode. The electrode surface of the invasive ECoG stereotactic deep brain electrode is covered with a 10μm dexamethasone sustained-release layer and a 50μm parylene insulating layer. The non-invasive dry electrode EEG cap adopts an Ag / AgCl sintered electrode and a conductive gel micro-reservoir structure with an impedance of <5kΩ.

[0008] b) Adaptive neural decoding algorithm module, which adopts CNN-LSTM-Transformer hybrid network architecture, recursively updates weights every 15s, and outputs 7-dimensional lower limb movement intention, which includes left / right hip flexion and extension, left / right knee flexion and extension, left / right ankle flexion and extension, and stillness;

[0009] c) A three-joint exoskeleton, which uses a layered control strategy to control the movement of the exoskeleton joints in real time according to the stated movement intention and dynamically adjusts the gait parameters;

[0010] d) Redundancy switching unit, which switches to full-function EEG mode within 200ms when ECoG signal is abnormal;

[0011] e) Spinal cord stimulation module, used to provide closed-loop auxiliary stimulation when the EEG signal is weak, including an electrode array placed outside the dura mater, outputting a stimulation current with a frequency of 40Hz and a pulse width of 300μs.

[0012] Preferably, the sampling rate of the ECoG electrode array is 1kHz, and a differential amplifier circuit is used to suppress metal artifacts; the signal transmission of the ECoG electrode array uses 13.56MHz wireless power supply and 402~405MHz UHF data transmission.

[0013] Preferably, the adaptive neural decoding algorithm module includes:

[0014] Time-frequency feature extraction: The β band of 13-30Hz and the γ band of 30-100Hz are separated by short-time Fourier transform;

[0015] Spatial feature enhancement: Co-space mode CSP filtering focuses on the C3 / C4 / Cz region;

[0016] Real-time calibration: Model parameters are updated every 15 seconds to compensate for signal drift.

[0017] Preferably, the hierarchical control strategy includes:

[0018] High-level control: converting movement intentions into joint angle trajectories (hip flexion 0-30°);

[0019] Mid-level control: Model predictive control (MPC) optimizes gait stability, with a prediction time of 0.5s;

[0020] Low-level control: PID motor tracking, proportional gain Kp=2.5, integral time Ti=0.1s.

[0021] Preferably, the dynamic adjustment of gait parameters further includes the use of a terrain adaptation mechanism, which includes: constructing an environmental map based on lidar and RGB-D camera; and adjusting stride length (±20%) and leg lift height (±10cm) in real time to cross obstacles.

[0022] Preferably, the process of constructing an environmental map based on LiDAR and RGB-D cameras, and adjusting stride length and leg lift height in real time to cross obstacles is as follows:

[0023] Step 1, Multi-sensor synchronous acquisition: The RGB-D camera outputs a 640×480 depth map at 30fps, and the 16-line LiDAR outputs a point cloud at 10Hz; the two data streams are synchronized on Jetson Xavier NX via PTP time synchronization to ensure that the time deviation is less than 1 millisecond;

[0024] Step 2, Point Cloud Preprocessing and Fusion: The original point cloud of the LiDAR is subjected to 2cm voxel filtering, and then the ground is segmented using RANSAC, retaining only the points within the range of 0-50cm above the ground; the RGB-D depth map is converted into a 3D point cloud after bilateral filtering, and then stitched with the LiDAR point cloud in a unified coordinate system to form a hybrid point cloud.

[0025] Step 3, Local 2.5D Raster Mapping: Project the mixed point cloud onto a 5cm×5cm 2D raster, storing the maximum height h_max in each raster; update the raster map at a frequency of 30Hz and maintain the sliding window of the most recent 2s.

[0026] Step 4, Obstacle Detection and Classification: Calculate the height difference Δh = h_max − h_ground for each grid cell. If Δh is greater than 2cm and three consecutive grid cells satisfy this condition, then the area is marked as an obstacle. Output obstacle box parameters: center coordinates (x,y), width w, depth d, and height h.

[0027] Step 5, real-time gait parameter decision-making: Calculate the leg lift height H_step=max(5cm,h+3cm) based on the obstacle height h, and ensure it does not exceed 40% of the user's leg length; adjust the stride length based on the obstacle width w: increase the stride length by 20% when w<10cm, increase it by 10% when w–20cm, and decrease it by 20% when w>20cm; the baseline stride length L0 is preset by the user; all parameters are filtered by a 2Hz low-pass filter, and if the height change between adjacent frames is >3cm, a 300ms Bezier curve is used for smooth transition;

[0028] Step 6, Trajectory Generation and Optimization: Input the adjusted H_step and L_step into the min-jerk trajectory generator to obtain the angle-time series of the hip, knee, and ankle joints; optimize joint torque within the 0.5s prediction domain using the mid-level model prediction control; and control the angle tracking error within ±1° using the bottom-level PID driver brushless motor.

[0029] Step 7, Safety and Anomaly Handling: If no overlap between RGB-D and LiDAR point clouds is detected for 5 consecutive frames, remove the obstacle marker; if the exoskeleton joint angular velocity exceeds 120° / s or the foot force is greater than 150N, immediately trigger an emergency stop; if the sensor fails, the system will automatically degrade, shorten the stride by 20%, and fix the leg lift height to 5cm.

[0030] Preferably, the neural function remodeling method includes:

[0031] Gradual disengagement: Reduce the assist ratio weekly, from 100% to 30%;

[0032] Hebbian Plasticity Training: Exoskeleton Response within 150ms after Motor Imagination Begins.

[0033] Preferably, it further includes:

[0034] Electromyography (EMG) monitoring module: Collects EMG data from the rectus femoris / tibialis anterior muscle and cuts off the output when the stimulation current exceeds the threshold.

[0035] Gamified rehabilitation interface: Virtual character movements are synchronized with the exoskeleton in real time.

[0036] Based on the above technical solution, the advantages of the present invention are:

[0037] (1) High-precision and high-reliability motion control:

[0038] This invention employs dual-modal signal acquisition (invasive ECoG stereotactic deep brain electrodes + non-invasive EEG cap), combining the high signal-to-noise ratio (25dB) of ECoG with the low impedance (<5kΩ) of EEG to achieve complementary advantages: ECoG provides millimeter-level spatial resolution (capable of distinguishing individual joint movements), while EEG provides redundancy. This solves the problem of insufficient decoding accuracy of EEG signals in existing technologies, improving the decoding accuracy of 7-dimensional lower limb movement intentions (left / right hip flexion / extension, left / right knee flexion / extension, left / right ankle flexion / extension, and stillness) to ≥95%, with a latency of <100ms.

[0039] The adaptive neural decoding algorithm module (CNN-LSTM-Transformer hybrid network architecture) recursively updates weights every 15 seconds, separates the β band (13-30Hz) and γ band (30-100Hz) through short-time Fourier transform, and uses common spatial pattern (CSP) filtering to focus on key brain regions (C3 / C4 / Cz) to compensate for signal drift, significantly improving real-time performance and robustness.

[0040] (2) Adaptive interaction capability in complex environments:

[0041] This invention employs a hierarchical control strategy (high-level control: movement intention and joint angle trajectory; mid-level control: model predictive control (MPC) to optimize gait stability; low-level control: PID motor tracking), combined with a terrain adaptation mechanism (real-time mapping using LiDAR and RGB-D cameras), to dynamically adjust stride length (±20%) and leg lift height (±10cm), effectively addressing complex terrains such as slopes and obstacles. This solves the problem of poor adaptability in existing exoskeletons, ensuring that the exoskeleton joint angle error is <±2° and the control accuracy is ≥97%.

[0042] When the ECoG signal is abnormal (such as impedance > 20kΩ or wireless packet loss rate > 5%), the redundant switching unit switches to the full-function EEG mode within 200ms to ensure uninterrupted system operation and improve clinical safety and reliability.

[0043] (3) Promote neurological remodeling and rehabilitation:

[0044] This invention strengthens the cortical-muscle pathway and promotes spinal cord plasticity through a progressive disengagement mechanism (reducing the assistance ratio weekly from 100% to 30%) and Hebbian plasticity training (the exoskeleton responds within 150ms after the start of motor imagery). This addresses the problems of long-term reliance on external devices and insufficient nerve recovery; clinical data show that patients' voluntary motor ability improves by 40% when the assistance ratio is reduced to 30%.

[0045] The spinal cord stimulation module provides closed-loop auxiliary stimulation (40Hz frequency, 300μs pulse width current) when the EEG signal is weak. Combined with the electromyography monitoring module (rectus femoris / tibialis anterior muscle EMG, output is cut off when stimulation exceeds the threshold), safe and closed-loop neurorehabilitation is achieved, avoiding overstimulation.

[0046] (4) Security and user experience optimization:

[0047] This invention integrates electromyography (EMG) monitoring and a gamified rehabilitation interface (real-time synchronization between the Unity3D virtual character and the exoskeleton, with a latency of ≤50ms), enhancing patient immersion and improving rehabilitation compliance through visual, auditory, and tactile feedback. Simultaneously, an EMG verification mechanism (freezing the exoskeleton when the EMG timing difference from the intention is >200ms) prevents accidental triggering and ensures clinical safety.

[0048] The three-jointed exoskeleton features a lightweight design (carbon fiber structure, weighing 4.2kg) and multiple safety mechanisms (mechanical stops + electronic limits + emergency stop button), making it suitable for patients with a height of 155-190cm and ensuring long-term wearing comfort. Attached Figure Description

[0049] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0050] Figure 1 This is a diagram of the overall system architecture.

[0051] Figure 2 The flowchart of the adaptive decoding algorithm module is shown below.

[0052] Figure 3 This is a block diagram of a hierarchical control strategy;

[0053] Figure 4 Flowchart for dynamic adjustment of terrain-adaptive gait parameters;

[0054] Figure 5 This is a timing diagram for redundancy switching. Detailed Implementation

[0055] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0056] This invention provides an adaptive closed-loop exoskeleton system based on a multimodal brain-computer interface, such as... Figure 1 As shown, it includes:

[0057] a) A multimodal signal acquisition unit, comprising an invasive ECoG stereotactic deep brain electrode implanted in the lower limb functional area of ​​the motor cortex and a non-invasive dry electrode EEG cap. The invasive ECoG stereotactic deep brain electrode is a cylindrical electrode. The electrode surface of the invasive ECoG stereotactic deep brain electrode is covered with a 10μm dexamethasone sustained-release layer and a 50μm parylene insulating layer. The non-invasive dry electrode EEG cap adopts an Ag / AgCl sintered electrode and a conductive gel micro-reservoir structure with an impedance of <5kΩ.

[0058] b) Adaptive neural decoding algorithm module, which adopts CNN-LSTM-Transformer hybrid network architecture, recursively updates weights every 15s, and outputs 7-dimensional lower limb movement intention, which includes left / right hip flexion and extension, left / right knee flexion and extension, left / right ankle flexion and extension, and stillness;

[0059] c) A three-joint exoskeleton, which uses a layered control strategy to control the movement of the exoskeleton joints in real time according to the stated movement intention and dynamically adjusts the gait parameters;

[0060] d) Redundancy switching unit, which switches to full-function EEG mode within 200ms when ECoG signal is abnormal;

[0061] e) Spinal cord stimulation module, used to provide closed-loop auxiliary stimulation when the EEG signal is weak, including an electrode array placed outside the dura mater, outputting a stimulation current with a frequency of 40Hz and a pulse width of 300μs.

[0062] The multimodal signal acquisition unit includes an invasive part and a non-invasive part.

[0063] Specifically, the invasive ECoG stereotactic deep brain electrode uses a 64-channel ECoG electrode array, with cylindrical electrodes, 2 mm in diameter and 4 mm apart, implanted in the lower limb representative area of ​​the anterior central gyrus (MRI-T1 navigation positioning). The electrode surface consists of a 10 μm dexamethasone sustained-release layer with a drug loading of 0.3 μg / mm², linear release over 90 days; and a 50 μm parylene-C insulating layer with a breakdown voltage of 200 V.

[0064] The cylindrical electrode designed in this invention is developed for stereotactic electroencephalography (SEEG). With a diameter of 2mm, its advantage lies in its ability to be precisely implanted into specific functional areas deep within the brain (such as the lower limb representation area of ​​the precentral gyrus) using a minimally invasive approach, directly recording electrical signals from deep neural nuclei with higher spatial resolution. The 4mm electrode spacing helps cover the target area and reduce crosstalk between signals. The electrode contacts can be made of a biocompatible platinum-iridium alloy, exhibiting stable electrochemical performance. The key insulating material, parylene-C, is a conformal coating material widely used in long-term implantable devices. Its 50μm thickness and 200V breakdown voltage ensure long-term insulation reliability and patient safety in the in vivo environment. The dexamethasone sustained-release layer is an important innovative design, aiming to effectively inhibit the formation of glial scar tissue around the implant by continuously releasing the anti-inflammatory drug, thereby ensuring long-term stable signal recording quality.

[0065] The sampling parameters are as follows: differential amplifier CMRR≥120dB, input noise<1μVRMS; wireless link: 13.56MHz inductive power supply, 402-405MHz UHF data backhaul, bandwidth 250kbps, delay<10ms.

[0066] The non-invasive dry electrode EEG cap has 32 channels, the electrode is an Ag / AgCl sintered electrode, and the electrode is equipped with a conductive gel micro reservoir with a specification of Φ3mm×depth2mm. The impedance of a single wear is <5kΩ and it can be used continuously for 8 hours.

[0067] The fundamental purpose of this invention's dual-modal design of invasive ECoG + non-invasive EEG is to achieve the optimal balance between engineering and clinical application between "high-fidelity neural control" and "long-term safety and usability." The specific reasons are explained below:

[0068] 1. Complementarity: Addressing the fatal shortcomings of single-modality architectures

[0069] Dimension Invasive ECoG Non-invasive EEG Dual-modal fusion benefits Signal quality High (μV level noise, bandwidth 0.5-300Hz) Low (mV level noise, bandwidth 0.5-40Hz) ECoG enables fine-grained control, while EEG provides redundancy verification. Spatial resolution Millimeter-level accuracy (can distinguish individual finger movements) Centimeter-level (can only distinguish large limb movements) Achieve independent control of multiple joints (such as ankle dorsiflexion + hip abduction). Long-term stability Signal attenuation <5% one year after implantation Electrode displacement causes signal drift >20% / month EEG serves as a fallback backup in case of ECoG failure. Surgical risks Brain surgery is required for implantation (infection rate 1-3%). Non-invasive (risk of skin irritation only) A phased treatment strategy is permitted (EEG verification first, then implantation).

[0070] 2. Triple Needs Driven by Clinical Scenarios

[0071] (1) Acute phase: EEG-based rapid initiation of rehabilitation

[0072] Within 7 days post-surgery: The patient's cerebral edema had not subsided, posing a risk to ECoG implantation. EEG was used first for bedside passive training to prevent muscle atrophy. Data shows that upgrading to ECoG implantation is only recommended when the EEG decoding accuracy is >75%.

[0073] (2) Chronic phase: ECoG-controlled fine-tuning

[0074] Walking on complex terrain: It is necessary to adjust the angles of 7 joints in real time (such as going up and down stairs). Only ECoG can meet the requirements of multi-degree-of-freedom synchronous control with a delay of <50ms.

[0075] Neuroplasticity training: The high-frequency gamma band (80-150Hz) of ECoG can precisely trigger Hebbian synaptic reinforcement.

[0076] (3) Emergency response: Seamless EEG switching

[0077] Implant infection / malfunction: When the ECoG signal is lost, the system automatically switches to EEG mode to maintain basic walking functions, such as walking on flat ground.

[0078] Therefore, the technical essence of adopting a dual-modal design is a clinical compromise between "high risk and high reward" and "low risk and universality": ECoG is responsible for "ultimate performance" (such as walking on complex terrain and neural pathway remodeling); EEG ensures "bottom-line safety" (such as emergency control after implantation failure).

[0079] Furthermore, the adaptive neural decoding algorithm module adopts a CNN-LSTM-Transformer hybrid network architecture, such as... Figure 2 As shown, it includes:

[0080] Time-frequency feature extraction: The β band of 13-30Hz and the γ band of 30-100Hz are separated by short-time Fourier transform;

[0081] Spatial feature enhancement: Co-space mode CSP filtering focuses on the C3 / C4 / Cz region;

[0082] Real-time calibration: Model parameters are updated every 15 seconds to compensate for signal drift.

[0083] Specifically, its time-frequency characteristics and spatial characteristics are set as follows:

[0084] Time-frequency characteristics: 256-point STFT, window length 200ms, 50% overlap, extracting β power spectrum from 13-30Hz and γ power spectrum from 30-100Hz.

[0085] Spatial characteristics: Common spatial mode (CSP) filter, selecting 8 channels in the C3 / C4 / Cz region, retaining the first 3 spatial filters.

[0086] Furthermore, the network structure of the CNN-LSTM-Transformer hybrid network architecture includes:

[0087] CNN: Two 1-D convolutional layers, kernel length 16 / 32, ReLU activation;

[0088] LSTM: 128 hidden units, dropout 0.2;

[0089] Transformer: 4-head attention, 64-dimensional embedding.

[0090] Training methods include: cross-entropy loss, Adam optimizer, and a learning rate of 10⁻. 4 The weights are updated online every 15 seconds (incremental learning, β=0.99).

[0091] The output includes a 7-dimensional probability vector consisting of L-Hip, R-Hip, L-Knee, R-Knee, L-Ankle, R-Ankle, and Rest, with a threshold of 0.8.

[0092] Furthermore, the mechanical structure of the three-joint exoskeleton is a carbon fiber thigh-lower leg linkage-hip / knee / ankle 3DOF structure with a peak torque of 80 Nm and a weight of 4.2 kg. Its sensor setup includes: a 14-bit joint encoder, a six-dimensional force sensor on the sole of the foot with a range of 0-200 N and an accuracy of 0.5 N; and a lidar with a range of 0.1-8 m and a resolution of 1°.

[0093] The structure includes active actuation of 3 degrees of freedom (DOF) for the hip, knee, and ankle, as well as one passive dorsiflexion / plantarflexion joint (ankle elastic hinge). The peak torque per joint is 80 N·m, and the entire device weighs ≤4.2 kg (including battery, controller, and sensors). It is suitable for heights of 155–190 cm; the range of motion meets rehabilitation needs for walking on flat ground, climbing stairs, navigating slopes, and overcoming obstacles.

[0094] Specifically, the overall configuration and degrees of freedom are as follows:

[0095] Hip: 1DOF (sagittal plane flexion and extension), the center of rotation is concentric with the femoral head.

[0096] Knee: 1DOF (flexion and extension), center of rotation concentric with the tibial plateau.

[0097] Ankle: 1DOF (dorsiflexion / plantarflexion), rotation center concentric with the talus; passive inversion / eversion spring hinges provide ±5° elastic degrees of freedom.

[0098] 1. Main Structure

[0099] Carbon fiber composite material: The thigh connecting rod is made of T700 unidirectional fabric and epoxy hot-pressed, with an outer diameter of 28mm, a wall thickness of 2mm, a bending stiffness of 580N·m², and a density of 1.6g / cm³. The lower leg connecting rod is made of the same material, with an outer diameter of 25mm and a wall thickness of 1.8mm.

[0100] Joint housing: CNC machined 7075-T6 aluminum alloy, anodized; the housing and carbon tube are connected by tapered adhesive bonding and rivets, with a torsional strength >150 N·m.

[0101] 2. Drive unit

[0102] Part Motor Model reducer peak torque Continuous torque weight Hip EC-4pole200W Harmonic 1:100 80 N·m 35 N·m 0.65kg knee EC-4pole 150W Harmonic 1:120 80 N·m 30 N·m 0.60kg ankle EC-4pole 100W Planet 1:40 80 N·m 25 N·m 0.45kg

[0103] The motor features a 14-bit absolute encoder and built-in temperature control (power reduction above 70℃). The harmonic reducer has zero backlash (<1 arcmin), ensuring closed-loop position accuracy.

[0104] 3. Lightweight accessories

[0105] Battery: 21700 lithium polymer 12V-3.2Ah, placed behind the waist, weight 0.38kg.

[0106] Controller: STM32H7 + CAN-FD bus, onboard IMU (BMI270), weight 0.12kg.

[0107] 4. Sensor Layout

[0108] 4.1 Joint Encoder

[0109] A 14-bit magnetoelectric absolute encoder with a resolution of 0.022°; it is installed coaxially with the rear end of the motor, and eccentricity error is eliminated through a flexible coupling.

[0110] 4.2 Foot six-dimensional force sensor

[0111] ATIMini45 or equivalent domestically made, measuring range 0-200N (Fz), accuracy ±0.5N; installed under the carbon fiber base plate of the footrest, with a 2mm TPU cushioning pad between it and the sole of the shoe for impact protection.

[0112] 4.3 LiDAR

[0113] RPLIDARA3M1, 0.1-8m distance, 1° angular resolution, 10Hz; mounted on the front of the lower leg, tilted downwards at 15°, scanning a 270° area in front; fused with an IMU, generating a 2.5D raster map through extended Kalman filter (EKF).

[0114] 4.4 Auxiliary Sensors

[0115] IMU: 6-axis waist measurement for attitude estimation; Temperature: NTC 10kΩ motor winding for real-time monitoring to prevent overheating; Pressure membrane: 4 points on the inside of the strap to detect tightness and prevent slippage.

[0116] 5. Joint mechanical restraint and safety

[0117] Mechanical stops are set as follows: Hip flexion / extension: -30° to +45°; Knee flexion / extension: -5° to +120°; Ankle dorsiflexion: -20° to +30°;

[0118] Electronic limit switch: The software limit switch triggers 5° earlier than the mechanical limit switch, with dual redundancy.

[0119] Emergency stop buttons are set up: one on each side of the left and right thighs. Pressing the button will cut off the motor power 10ms later.

[0120] 6. Size and Fit

[0121] Adjustable thigh length: Utilizes a carbon fiber telescopic sleeve and quick-release buckle design, adjustable range 35-45cm. Adjustable calf length: Same structure as thigh, 30-40cm. Foot support size: S / M / L, secured with Velcro.

[0122] Furthermore, the three-joint exoskeleton uses a layered control strategy to control the exoskeleton joint movements in real time, following the stated movement intention, and dynamically adjusts gait parameters. For example... Figure 3 As shown, the hierarchical control strategy includes:

[0123] High-level control: 7-dimensional intent—joint angle trajectory (e.g., hip flexion 0-30°), using minimum jerk planning.

[0124] Mid-level control: MPC prediction time domain 0.5s, sampling period 20ms, cost function J=Σ(attitude error²+torque²).

[0125] Low-level control: PID parameters Kp=2.5, Ki=25, Kd=0.05, current loop 1kHz, integral time Ti=0.1s.

[0126] The dynamic adjustment of gait parameters also includes the use of a terrain adaptation mechanism, which includes: real-time mapping using an RGB-D camera (30fps) and LiDAR, and the algorithm dynamically adjusts the stride length by ±20% and the leg lift height by ±10cm based on the obstacle height.

[0127] Specifically, the dynamic adjustment of terrain-adaptive gait parameters includes:

[0128] 1. Hardware Components

[0129] RGB-D camera: Microsoft Azure Kinect DK, 30fps, 640×480 depth map, effective range 0.1-8m, depth error <5mm.

[0130] LiDAR: 16-line LiDAR, horizontal angular resolution 0.2°, vertical -15°~+15°, 10Hz point cloud output.

[0131] Fusion board: NVIDIA Jetson Xavier NX, 8GB RAM, Ubuntu 20.04 + ROS2Foxy.

[0132] 2. Coordinate System and Calibration

[0133] Exoskeleton base coordinate system O e With the center of the pelvis as the origin, X points forward, Y points to the left, and Z points upward. Camera coordinate system O. c The rigid transformation T is obtained through hand-eye calibration with the LiDAR coordinate system Oʟ. ce T Le The calibration error is <1cm. All sensors are synchronized via PTP time, with a time deviation of <1ms.

[0134] 3. Real-time mapping process

[0135] Point cloud preprocessing

[0136] The original LiDAR point cloud was subjected to voxel filtering (2cm voxel side length) and ground segmentation (RANSAC plane fitting). The remaining point cloud only retained the area 0-25cm above the ground (small obstacles) and 25-50cm above the ground (large obstacles). The RGB-D depth map was then converted into a 3D point cloud through bilateral filtering and depth thresholding, and stitched together with the LiDAR point cloud in the same coordinate system.

[0137] Local raster map

[0138] A 2.5D elevation grid is used, with a grid size of 5cm × 5cm, and each cell stores the maximum height h. max Update frequency: 30Hz (synchronized with RGB-D frames), the sliding window keeps the most recent 2 seconds of data to prevent drift.

[0139] Obstacle detection

[0140] Calculate the height difference Δh=h for each grid cell. max -h 9roun If Δh > 2cm and three consecutive grid cells satisfy this condition, then mark it as an obstacle; output the obstacle bounding box: center (x,y), width w, depth d, height h.

[0141] 4. Gait parameter decision algorithm

[0142] Input: obstacle box height h; current gait cycle percentage φ∈[0,1], calculated by plantar force sensor; patient height H (cm) and leg length L le9 (cm).

[0143] Adjustment rules:

[0144] a) Leg lift height H step H step =max(5cm,h+3cm); Constraint: H step ≤0.4·L le9 (Prevent falls).

[0145] b) Step size L step Obstacle width w < 10cm: L step =L0·(1+0.2); where, 10cm≤w≤20cm: L step =L0·(1+0.1); w>20cm:L step =L0·(1−0.2); where L0 is the baseline step size set by the current user.

[0146] Online smoothing:

[0147] A first-order low-pass filter (cutoff frequency 2Hz) is used to avoid abrupt changes; if an obstacle height change of >3cm is detected between two consecutive frames, second-order Bezier curve interpolation is enabled with a transition time of 300ms.

[0148] 5. Exoskeleton execution

[0149] High-level trajectory: (H) step ,L step Input the min-jerk trajectory generator and output the timing of hip, knee, and ankle angles; the middle layer MPC optimizes joint torque within the 0.5s prediction domain to ensure that the sole of the foot does not collide with obstacles; the bottom layer PID drives the brushless motor with an angle tracking error ≤ ±1°.

[0150] 6. Safety and Abnormal Handling

[0151] Obstacle misjudgment: If no RGB-D depth data overlaps with the LiDAR point cloud for 5 consecutive frames, the obstacle marker is removed;

[0152] Emergency Stop: Stop the machine immediately when the exoskeleton joint angular velocity is >120° / s or the plantar force is >150N;

[0153] Degradation mode: When the sensor fails, the default step length is reduced by 20% and the leg lift height is fixed at 5cm.

[0154] Specifically, such as Figure 4 As shown, the process of building an environmental map based on LiDAR and RGB-D cameras, and adjusting stride length and leg lift height in real time to cross obstacles is as follows:

[0155] Step 1, Simultaneous data acquisition by multiple sensors:

[0156] The RGB-D camera outputs a 640×480 depth map at 30fps, and the 16-line LiDAR outputs a point cloud at 10Hz; the two data streams are synchronized on the Jetson Xavier NX via PTP time synchronization to ensure a time deviation of less than 1 millisecond.

[0157] Step 2, Point Cloud Preprocessing and Fusion:

[0158] The original point cloud of the lidar is filtered by 2cm voxels, and then the ground is segmented using RANSAC, retaining only the points within the range of 0–50cm above the ground; the RGB-D depth map is converted into a 3D point cloud after bilateral filtering, and then stitched with the lidar point cloud in a unified coordinate system to form a hybrid point cloud.

[0159] Step 3, Local 2.5D raster mapping:

[0160] The mixed point cloud is projected onto a 5cm×5cm 2D grid, and each grid stores the maximum height h_max; the grid map is updated at a frequency of 30Hz, and the sliding window of the most recent 2s is maintained to prevent dynamic objects from drifting.

[0161] Step 4, Obstacle Detection and Classification:

[0162] Calculate the height difference Δh = h_max − h_ground for each grid cell. If Δh is greater than 2cm and three consecutive grid cells meet this condition, then mark the area as an obstacle. Output obstacle box parameters: center coordinates (x,y), width w, depth d, and height h.

[0163] Step 5, Real-time decision-making on gait parameters:

[0164] The leg lift height H_step is calculated based on the obstacle height h: H_step = max(5cm, h + 3cm), and must not exceed 40% of the user's leg length. The step size is adjusted based on the obstacle width w: increase the step size by 20% when w < 10cm, increase it by 10% when w–20cm, and decrease it by 20% when w > 20cm. The baseline step size L0 is preset by the user. All parameters are low-pass filtered at 2Hz. If the height change between adjacent frames is greater than 3cm, a 300ms Bezier curve is used for smooth transition.

[0165] Step 6, Trajectory Generation and Optimization:

[0166] The adjusted H_step and L_step are input into the min-jerk trajectory generator to obtain the angle-time series of the hip, knee and ankle joints; the middle-level model predictive control (MPC) optimizes the joint torque within the 0.5s prediction domain to ensure that the sole of the foot does not collide with the obstacle; the bottom-level PID drives the brushless motor, and the angle tracking error is controlled within ±1°.

[0167] Step 7, Safety and Abnormal Handling:

[0168] If no overlap between RGB-D and LiDAR point clouds is detected for 5 consecutive frames, the obstacle marker is removed; if the angular velocity of the exoskeleton joint exceeds 120° / s or the foot force is greater than 150N, an emergency stop is triggered immediately; in case of sensor failure, the system automatically degrades, shortens the stride by 20%, and fixes the leg lift height to 5cm to ensure user safety.

[0169] In this invention, the real-time control accuracy of the exoskeleton is ≥97% (7-dimensional joint angle error <±2°), which is achieved through a three-level closed loop of "signal-algorithm-execution" and dual correction of "multi-source feedback-predictive compensation". The specific technical path is as follows:

[0170] 1. Signal Terminal: Dual-mode high-precision intent acquisition

[0171] index Invasive ECoG Non-invasive EEG Synergistic gain Spatial resolution 2mm level (distinguishing by a single finger) 2cm level (limb differentiation) ECoG-led fine joint control Signal-to-noise ratio 25dB 12dB EEG as a drift correction reference Time delay <10ms (wireless link) <50ms (Bluetooth) Dual-channel parallel weighted fusion

[0172] Real-time calibration: The weights of the CNN-LSTM-Transformer model are updated every 15 seconds to compensate for electrode drift.

[0173] 2. Algorithm-side: Millisecond-level prediction - correction loop

[0174] 2.1 Hierarchical Control Architecture

[0175] EEG decoding output (7-dimensional) - trajectory planning (min-jerk) - MPC prediction (0.5s time domain) - PID motor tracking;

[0176] Trajectory planning: The minimum jerk model is adopted to ensure smooth joint trajectory and acceleration abrupt change <500° / s², avoiding mechanical oscillation.

[0177] MPC model prediction: State equation: x(k+1)=Ax(k)+Bu(k), where A includes exoskeleton inertia and damping parameters;

[0178] Rolling optimization: Solve the quadratic programming every 20ms to predict the joint angle error for the next 10 steps to be <±1°.

[0179] PID disturbance rejection: Parameters: Kp=2.5, Ki=25, Kd=0.05 (tuned by Ziegler-Nichols method);

[0180] Feedforward compensation: By incorporating gravitational torque and frictional force feedforward, the error is reduced by 30%.

[0181] 2.2 Dynamic Weight Fusion

[0182] Signal confidence calculation:

[0183] Confidence level = 0.9 × ECoG_SNR / (ECoG_SNR + EEG_SNR) + 0.1 × historical accuracy: When ECoG signal-to-noise ratio > 20dB, ECoG weight > 90% to ensure high accuracy; when ECoG signal-to-noise ratio < 15dB, EEG weight is increased to 50% to avoid control interruption.

[0184] 3. Execution end: Employs microsecond-level feedback correction.

[0185] 3.1 Multi-source sensor fusion, the specific fusion method is shown in the table below:

[0186] sensor accuracy effect Joint encoder 0.1° Real-time angle tracking Foot six-dimensional force sensor 0.5N Ground reaction force—corrected joint torque IMU (lumbar region) 0.5° Posture prediction—preventing falls

[0187] Extended Kalman Filter (EKF) is used: encoder and IMU data are fused to estimate joint angular velocity error < ±0.5° / s.

[0188] 3.2 Motor-level closed loop

[0189] On the current loop: 1kHz PID control, current tracking error <±1% of rated value;

[0190] In terms of feedforward friction compensation: the Lu-Gre friction model is adopted, and the error is reduced by 40% at low speeds (<5° / s).

[0191] 4. Fault tolerance mechanism

[0192] This invention employs EMG verification: if the timing difference between the EMG and the decoding intent is >200ms, the exoskeleton is forcibly frozen to prevent false triggering;

[0193] Redundancy switching: In the event of an ECoG failure, the EEG single-mode still maintains an accuracy of ±3° and a delay of <200ms.

[0194] Therefore, this invention achieves real-time 7-dimensional joint control error of <±2° and delay of <100ms by superimposing dynamic weight fusion and EKF sensor fusion through a three-level closed loop of "μV-level EEG signal - millisecond-level algorithm prediction - microsecond-level motor response". This meets the high-precision rehabilitation needs in complex terrain.

[0195] Furthermore, the redundancy switching unit switches to full-function EEG mode within 200ms when the ECoG signal is abnormal. The triggering condition can be set to: ECoG channel impedance > 20kΩ or wireless packet loss rate > 5%. Specifically, its switching logic is: shut down the ECoG decoding thread and start the EEG thread within 200ms to keep the exoskeleton running without interruption.

[0196] like Figure 5 As shown, the workflow steps of the redundancy switching unit are as follows:

[0197] 1) The monitoring thread collects ECoG channel impedance and wireless packet loss rate every 10ms;

[0198] 2) If the impedance is greater than 20kΩ or the packet loss rate is greater than 5%, it will be judged as abnormal after 3 consecutive occurrences, and a 200ms countdown will be started.

[0199] 3) During the countdown: Freeze the ECoG decoding thread output and save the current joint angle as the initial EEG value;

[0200] 4) Within 200ms: shut down the ECoG thread, start the EEG thread, increase the weight from 0 to 100% and perform smooth interpolation.

[0201] 5) The switching completion signal is fed back to the main controller, and the exoskeleton continues to run without any sudden angle changes.

[0202] Furthermore, to provide closed-loop auxiliary stimulation when EEG signals are weak, the system of this invention also includes a spinal cord stimulation module, employing 16-contact paddle electrodes with contacts measuring 1.5mm × 4mm and a spacing of 1mm × 4.5mm, implanted epidurally at T11-L2. The electrode array parameters are set as follows: 40Hz, pulse width 300μs, current 0-25mA, and closed-loop threshold: activated when the 7-dimensional intent confidence level is <0.5 and the EMG amplitude is <10μV.

[0203] For enhanced system safety, an electromyography (EMG) monitoring module is also included. Specifically, EMG acquisition includes wireless surface electrodes for the rectus femoris and tibialis anterior muscles, with a sampling rate of 1 kHz. If the stimulation-induced EMG RMS > 100 μV, the stimulation current is immediately cut off.

[0204] To further meet the needs of neural functional remodeling, the system incorporates a gamified interface. When using the system, a gradual disengagement approach is adopted, such as a 4-week cycle where the assistance ratio decreases by 15% each week, eventually reaching 30%. The specific settings can be flexibly adjusted according to the patient's actual situation. Simultaneously, Hebbian training can be performed, where the exoskeleton provides assistance within 150ms after detecting motor imagery, strengthening the cortical-muscle pathway.

[0205] The gamification interface is based on a Unity3D scene, with virtual characters and exoskeletons synchronized in real time, providing visual, auditory, and tactile feedback. Specifically, it is implemented through the following "three-layer architecture and dual-channel closed loop":

[0206] Perception layer: exoskeleton + multiple sensors to obtain joint angle / foot force / terrain information;

[0207] Transport layer: 90Hz low-latency UDP—Unity3D scene;

[0208] Presentation layer: Unity3D humanoid avatar + visual, auditory, and tactile feedback - patient immersion training;

[0209] Target: End-to-end delay ≤ 50ms, joint angle synchronization error ≤ ±2°.

[0210] Compared with traditional technical solutions, the technical effects achieved by the above-mentioned technical solution of the present invention are as follows:

[0211] (1) High-precision and high-reliability motion control:

[0212] This invention employs dual-modal signal acquisition (invasive ECoG stereotactic deep brain electrodes + non-invasive EEG cap), combining the high signal-to-noise ratio (25dB) of ECoG with the low impedance (<5kΩ) of EEG to achieve complementary advantages: ECoG provides millimeter-level spatial resolution (capable of distinguishing individual joint movements), while EEG provides redundancy. This solves the problem of insufficient decoding accuracy of EEG signals in existing technologies, improving the decoding accuracy of 7-dimensional lower limb movement intentions (left / right hip flexion / extension, left / right knee flexion / extension, left / right ankle flexion / extension, and stillness) to ≥95%, with a latency of <100ms.

[0213] The adaptive neural decoding algorithm module (CNN-LSTM-Transformer hybrid network architecture) recursively updates weights every 15 seconds, separates the β band (13-30Hz) and γ band (30-100Hz) through short-time Fourier transform, and uses common spatial pattern (CSP) filtering to focus on key brain regions (C3 / C4 / Cz) to compensate for signal drift, significantly improving real-time performance and robustness.

[0214] (2) Adaptive interaction capability in complex environments:

[0215] This invention employs a hierarchical control strategy (high-level control: movement intention and joint angle trajectory; mid-level control: model predictive control (MPC) to optimize gait stability; low-level control: PID motor tracking), combined with a terrain adaptation mechanism (real-time mapping using LiDAR and RGB-D cameras), to dynamically adjust stride length (±20%) and leg lift height (±10cm), effectively addressing complex terrains such as slopes and obstacles. This solves the problem of poor adaptability in existing exoskeletons, ensuring that the exoskeleton joint angle error is <±2° and the control accuracy is ≥97%.

[0216] When the ECoG signal is abnormal (such as impedance > 20kΩ or wireless packet loss rate > 5%), the redundant switching unit switches to the full-function EEG mode within 200ms to ensure uninterrupted system operation and improve clinical safety and reliability.

[0217] (3) Promote neurological remodeling and rehabilitation:

[0218] This invention strengthens the cortical-muscle pathway and promotes spinal cord plasticity through a progressive disengagement mechanism (reducing the assistance ratio weekly from 100% to 30%) and Hebbian plasticity training (the exoskeleton responds within 150ms after the start of motor imagery). This addresses the problems of long-term reliance on external devices and insufficient nerve recovery; clinical data show that patients' voluntary motor ability improves by 40% when the assistance ratio is reduced to 30%.

[0219] The spinal cord stimulation module provides closed-loop auxiliary stimulation (40Hz frequency, 300μs pulse width current) when the EEG signal is weak. Combined with the electromyography monitoring module (rectus femoris / tibialis anterior muscle EMG, output is cut off when stimulation exceeds the threshold), safe and closed-loop neurorehabilitation is achieved, avoiding overstimulation.

[0220] (4) Security and user experience optimization:

[0221] This invention integrates electromyography (EMG) monitoring and a gamified rehabilitation interface (real-time synchronization between the Unity3D virtual character and the exoskeleton, with a latency of ≤50ms), enhancing patient immersion and improving rehabilitation compliance through visual, auditory, and tactile feedback. Simultaneously, an EMG verification mechanism (freezing the exoskeleton when the EMG timing difference from the intention is >200ms) prevents accidental triggering and ensures clinical safety.

[0222] The three-jointed exoskeleton features a lightweight design (carbon fiber structure, weighing 4.2kg) and multiple safety mechanisms (mechanical stops + electronic limits + emergency stop button), making it suitable for patients with a height of 155-190cm and ensuring long-term wearing comfort.

[0223] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them; although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications can still be made to the specific implementation of the present invention or equivalent substitutions can be made to some technical features without departing from the spirit of the technical solutions of the present invention, and all such modifications and substitutions should be covered within the scope of the technical solutions claimed in the present invention.

Claims

1. An adaptive closed-loop exoskeleton system based on a multimodal brain-computer interface, characterized in that: include: a) A multimodal signal acquisition unit, comprising an invasive ECoG stereotactic deep brain electrode implanted in the lower limb functional area of ​​the motor cortex and a non-invasive dry electrode EEG cap. The invasive ECoG stereotactic deep brain electrode is a cylindrical electrode. The electrode surface of the invasive ECoG stereotactic deep brain electrode is covered with a 10μm dexamethasone sustained-release layer and a 50μm parylene insulating layer. The non-invasive dry electrode EEG cap adopts an Ag / AgCl sintered electrode and a conductive gel micro-reservoir structure with an impedance of <5kΩ. b) Adaptive neural decoding algorithm module, which adopts CNN-LSTM-Transformer hybrid network architecture, recursively updates weights every 15s, and outputs 7-dimensional lower limb movement intention, which includes left / right hip flexion and extension, left / right knee flexion and extension, left / right ankle flexion and extension, and stillness; c) A three-joint exoskeleton, which uses a layered control strategy to control the movement of the exoskeleton joints in real time following the stated movement intention and dynamically adjusts gait parameters; d) Redundancy switching unit, which switches to full-function EEG mode within 200ms when ECoG signal is abnormal; e) Spinal cord stimulation module, used to provide closed-loop auxiliary stimulation when the EEG signal is weak, including an electrode array placed outside the dura mater, outputting a stimulation current with a frequency of 40Hz and a pulse width of 300μs.

2. The adaptive closed-loop exoskeleton system according to claim 1, characterized in that: The ECoG electrode array has a sampling rate of 1kHz and uses a differential amplifier circuit to suppress metal artifacts; the signal transmission of the ECoG electrode array uses 13.56MHz wireless power supply and 402~405MHz UHF data transmission.

3. The adaptive closed-loop exoskeleton system according to claim 1, characterized in that: The adaptive neural decoding algorithm module includes: Time-frequency feature extraction: The β band of 13-30Hz and the γ band of 30-100Hz are separated by short-time Fourier transform; Spatial feature enhancement: Co-space mode CSP filtering focuses on the C3 / C4 / Cz region; Real-time calibration: Model parameters are updated every 15 seconds to compensate for signal drift.

4. The adaptive closed-loop exoskeleton system according to claim 1, characterized in that: The hierarchical control strategy includes: High-level control: converting movement intentions into joint angle trajectories, where hip flexion is between 0-30°; Mid-level control: Model predictive control (MPC) optimizes gait stability, with a prediction time of 0.5s, a sampling period of 20ms, and a cost function J=Σ(attitude error²+torque²). Low-level control: PID motor tracking, proportional gain Kp=2.5, integral time Ti=0.1s.

5. The adaptive closed-loop exoskeleton system according to claim 1, characterized in that: The dynamic adjustment of gait parameters also includes the use of a terrain adaptation mechanism, which includes: An environmental map is built using LiDAR and an RGB-D camera; stride length and leg lift height are adjusted in real time to cross obstacles.

6. The adaptive closed-loop exoskeleton system according to claim 1, characterized in that: The neural function remodeling method includes: Gradual disengagement: Reduce the assist ratio weekly, starting from 100% and decreasing by 15% each week until the assist ratio reaches 30%; Hebbian Plasticity Training: Exoskeleton Response within 150ms after Motor Imagination Begins.

7. The adaptive closed-loop exoskeleton system according to claim 1, characterized in that: Also includes: Electromyography (EMG) monitoring module: Collects EMG data from the rectus femoris / tibialis anterior muscle and cuts off the output when the stimulation current exceeds the threshold.

8. The adaptive closed-loop exoskeleton system according to claim 1, characterized in that: Also includes: Gamified rehabilitation interface, virtual character movements are synchronized with exoskeleton in real time.

9. The adaptive closed-loop exoskeleton system according to claim 5, characterized in that: The process of using LiDAR and RGB-D cameras to build an environmental map and adjust stride length and leg lift height in real time to overcome obstacles is as follows: Step 1, Multi-sensor synchronous acquisition: The RGB-D camera outputs a 640×480 depth map at 30fps, and the 16-line LiDAR outputs a point cloud at 10Hz; the two data streams are synchronized on the Jetson Xavier NX via PTP time synchronization to ensure that the time deviation is less than 1 millisecond; Step 2, Point Cloud Preprocessing and Fusion: The original point cloud of the LiDAR is subjected to 2cm voxel filtering, and then the ground is segmented using RANSAC, retaining only the points within the range of 0-50cm above the ground; the RGB-D depth map is converted into a 3D point cloud after bilateral filtering, and then stitched with the LiDAR point cloud in a unified coordinate system to form a hybrid point cloud. Step 3, Local 2.5D Raster Mapping: Project the mixed point cloud onto a 5cm×5cm 2D raster, storing the maximum height h_max in each raster; update the raster map at a frequency of 30Hz and maintain the sliding window of the most recent 2s. Step 4, Obstacle Detection and Classification: Calculate the height difference Δh = h_max − h_ground for each grid cell. If Δh is greater than 2cm and three consecutive grid cells satisfy this condition, then the area is marked as an obstacle. Output obstacle box parameters: center coordinates (x,y), width w, depth d, and height h. Step 5, real-time gait parameter decision-making: Calculate the leg lift height H_step=max(5cm,h+3cm) based on the obstacle height h, and ensure it does not exceed 40% of the user's leg length; adjust the stride length based on the obstacle width w: increase the stride length by 20% when w<10cm, increase it by 10% when w–20cm, and decrease it by 20% when w>20cm; the baseline stride length L0 is preset by the user; all parameters are filtered by a 2Hz low-pass filter, and if the height change between adjacent frames is >3cm, a 300ms Bezier curve is used for smooth transition; Step 6, Trajectory Generation and Optimization: Input the adjusted H_step and L_step into the min-jerk trajectory generator to obtain the angle-time series of the hip, knee, and ankle joints; optimize joint torque within the 0.5s prediction domain using the mid-level model prediction control; and control the angle tracking error within ±1° using the bottom-level PID driver brushless motor. Step 7, Safety and Anomaly Handling: If no overlap between RGB-D and LiDAR point clouds is detected for 5 consecutive frames, remove the obstacle marker; if the exoskeleton joint angular velocity exceeds 120° / s or the foot force is greater than 150N, immediately trigger an emergency stop; if the sensor fails, the system will automatically degrade, shorten the stride by 20%, and fix the leg lift height to 5cm.

Images