*2.2. Movement Intention Recognition*

For patients with impaired motor function but can still move and do some daily activity, the willingness to actively participate in gait training is especially important in the rehabilitation process [37]. Clinical studies have shown that active involvement of patients in the rehabilitation training is more effective in the neurological reconstruction and motor function recovery. Therefore, as an important input information of active intelligent rehabilitation and assistant robotics device, human lower limb movement intention needs to be captured in real time.

The current typical intention recognition methods used in gait training systems and the statistics of recent research studies using each method are shown in Table 2. The neuro-rehabilitation training robotics devices should show "transparency" in the patient's walking assistance process, that is, reduce the intervention of the patient's active gait as much as possible [38], and the key lies in the understanding and prediction of the patient's movement intention. Current intention recognition methods are mainly based on bioelectric signals and motion signals. The electroencephalogram (EEG) signals are the overall reflection of the movement intentions in brain [39], and have the shortest latency, but it has a low signal-to-noise ratio, easily interfered by noise [40,41]. The Electromyographic (EMG) signals reflect the state of muscle activation and the feedback control based on EMG signal can effectively improve the human–machine coordination in rehabilitation training [42–45], but it has strong ambiguity and is affected by factors such as surface contact status, muscle displacement, and muscle fatigue [46]. The intention recognition method based on motion signal mainly uses kinematic signals such as position, angle, and speed, and kinetic signals such as interaction force/torque, which has high reliability, robustness, and accuracy [47–50]. Xu et al. [51] proposed a compliance control algorithm for walking-aid robots based on multi-sensor fusion, which allows the robot to obey human movement by recognizing user intentions. Esteban et al. [52] also carried out related research, using EMG signals and Artificial Neural Network (ANN) algorithms to recognize human walking intention and proposed a robotic knee exoskeleton for assistance and rehabilitation. Wu et al. [53] put forward a coordinated control strategy based on human– machine interaction and the principle of minimum interference. However, the information of human motion and force is the result of the movement, with a certain time lag between the motion intent. Therefore, in response to the active adjustment needs of human-inthe-loop control, it is necessary to study new motion perception systems and intention prediction models with self-learning capabilities, and to improve the stability, synergy and adaptability of human–machine collaboration based on active intention feedback.

**Table 2.** Intent recognition method used in each of the included studies and their characteristics.



**Table 2.** *Cont.*

## **3. Evaluation of Gait Rehabilitation**

Clinical gait analysis and evaluation is of great significance in active intelligent gait training systems. Quantitative analysis methods based on sensor data are important methods for gait rehabilitation evaluation. An increasing number of researchers in physical therapy, bioengineering, neurology and rehabilitation have been participating in this field of study. In the early research studies, gait analysis and evaluation usually took the form of scales, such as the Fugl-Meyer exercise scale [54]. According to the scale, medical staff perform the diagnosis and evaluation of motor function, the monitoring of disease progression, and the evaluation of curative effect. The result of evaluation is often affected by a large number of subjective and inaccurately measurable parameters in the clinical scale [55].

Table 3 shows number of gait evaluation studies, which sensors and features were used in each research [56–77], and the real-time of gait evaluation. Gait parameters are usually used to assist medical staff in diagnosis, rating and scoring of motor function, monitoring the progress of the patient's condition, and evaluating curative effect. Gait measurement equipment such as motion capture systems and wearable inertial sensors have been widely used in clinical practice. Some researchers used the gait parameters measured by these large systems to predict Parkinson's diagnosis and Hoehn-Yahr (H-Y) classification [67,68]. There are also researchers who used the changes in gait parameters before and after the patient receives treatment and training to evaluate the treatment effect [69]. Caramia et al. [70] used eight inertial measurement units placed on the lower extremities and trunk to estimate several gait parameters such as step length, stride speed, etc. and extract features from them to distinguish between healthy people and patients with H-Y grades 1 to 3 in order to achieve diagnosis and grade prediction. However, there are problems such as inconvenient use of sensing equipment, lack of clinical significance of data features, difficulty in matching the scale, and an incomplete assessing system. Wang et al. [71] carried out preliminary research based on clinical needs, using as few human sensor measurement data as possible, and using nonlinear data classification methods to achieve quantitative evaluation of dyskinesias in patients with abnormal gait. Skvortsov et al. [72] also investigated the feasibility of gait analysis and walking function evaluation based on the stance phase of stroke patients using biofeedback technology.

Muscle synergy theory describes a potential neuromuscular control mechanism of vertebrate limb movement [73]. According to the muscle synergy theory, nerves do not control a certain muscle alone, but recruit muscles on the spinal cord to form muscle groups, that is, muscle synergy. The muscles in the same muscle synergy are activated at the same time. Compared with controlling each muscle individually, using one control signal to activate multiple muscles theoretically provides a simplified system. Numerous experimental research results support this theory [74,75]. Studies have shown that muscle activation during motor tasks can be described in terms of low-dimensional control that reflects muscle synergy. The downward commands of the nervous system to the musculoskeletal system are manifested in muscle synergy, which is reflected in muscle activation

through spinal cord circuits or reflexes, thereby forming a force in the musculoskeletal system, driving the musculoskeletal system to move and producing specific actions.

**Table 3.** Sensors and features used in gait evaluation method in each of the included studies.


Abbreviations: + real-time evaluation; − off-line evaluation.

Based on the muscle synergy theory, many research studies have been carried out to diagnose gait disorders and neurological diseases by measuring the activation state of lower limb muscles during walking [76–80]. However, the existing methods for measuring muscles exercise have drawbacks. On the one hand, Surface Electromyography (sEMG) signal measurement has limitations which include the lack of ability to test the deep muscles, the easily interfered EMG sensors, and the difficulty for the extraction process of the EMG signal envelope to accurately demodulate the neural excitation when the motor neuron action potential is generated. On the other hand, Indwelling Electromyography (iEMG) causes a certain degree of damage to human muscles, which is not suitable for long-term exercise detection with multiple measurements. At the same time, the existing simulation software is generally based on a variety of rule constraints such as muscle forcelength relationship constraints, muscle force and joint motion coupling constraints, etc., and optimization theories such as minimizing physiological consumption. However, according to the results of human motion modeling and analysis by related researchers [79,80], in patients with gait disorders, it is often difficult to meet the above constraints due to nervemuscle-skeletal damage, and the dynamic representations such as joint torque are affected by motion compensation under the condition of external load changes.
