**5. Limitations and Challenges**

Rehabilitation and training of gait is a current research hot spot. From the systematic analysis of the current research status of the active intelligent gait training systems, it is not difficult to see that there are still key issues in terms of sensing, evaluation, and control. Key technologies such as the decoupling of radio frequency signals of human–machine coordinated movement, the understanding based on fusing physical models and gait data, and the generalized measurement of abnormal gait are in urgent need of breakthroughs. To be specific, when capturing patients' motion using RF signals, both the wearable gait training device and the human body reflect RF signals. That makes the decoupling of the

return signals from the two very important, and it is also the limitation of current research studies. The methods of lower limb movement analysis and movement intention prediction based on radio frequency principles need to be further studied. As the current intention recognition based on EEG signals is easily interfered by noise, the EMG signals-based method has strong ambiguity. Moreover, the intention recognition based on kinematic or kinetic signals has a long latency.

As a mobility aid for gait rehabilitation and training, if the evaluation criteria for the rehabilitation effect are difficult to define and the efficacy cannot be guaranteed, it will be difficult to meet the diverse and personalized needs of users. The current clinical scale for gait analysis and evaluation relies on the subjective assessment of the doctor and the self-perception of the patient. In addition, the existing sensing data features lack clinical significance and are difficult to correspond to the scale, and the evaluation system is inadequate. How to quantitatively evaluate the effect of gait training with multi-dimensional information still needs in-depth research.

A crucial problem of control of the active intelligent rehabilitation assistance devices for lower limbs is that it needs to allow the users to spontaneously participate in motion, which is of great importance for patients with nerve injury. However, the current gait rehabilitation training systems have difficulty accurately recognizing the user's movement intentions to make corresponding assistance strategies. As users' requirements for comfort and safety continue to increase, the human-in-the-loop control, with information of human body taking part in both the input and feedback of the controller, is receiving increasing attention. However, due to the difficulty of quickly and accurately identifying the user's intent, the research studies on human–machine cooperative intelligent control for personalized gait rehabilitation training is still too preliminary.

## **6. Future Directions**

To address the core problems in the sensing, evaluation, and control technology of the active intelligent gait training systems, the following possible future research directions are proposed. We believe that the key is to focus on scientific issues such as the decoding of lower limb movement intention based on the principle of radio frequency, the construction of a walking ability evaluation model combined with clinical knowledge base and lower limb movement data, and the personalized gait training strategy for collaborative control of human–machine systems.

Among the many methods of detecting and sensing human lower limb movement, the method based on the radio frequency signal is relatively preliminary, but it has obvious advantages and broad prospects. A new type of non-contact motion sensing method based on the principle of millimeter wave echo reflection needs to be studied. For instance, a non-contact small radio frequency sensor such as a millimeter wave radar first needs to be developed. Using the signal features generated by human motion on the radar image and Doppler signal spectrogram as target features, similar to Daniel et al. [110], and using the space occupancy status and motion frequency shift information contained in the frequency characteristic data of the range view as input, the features in the input data are encoded by the convolutional neural network (CNN) method, and an estimator is generated to output the joint position and motion information of the object [111]. By combining the real-time data with the models of the kinematics and dynamics of human lower limbs, the human lower limb movement may be predicted. In conclusion, it is of important scientific significance to study a new non-contact sensing principle and the method of model-driven and data-driven fusion, to integrate the characteristics of different information dimensions, to build a more concise, fast, and accurate online decoding model of composite information for patient's gait training, and to predict patient's movement intentions.

Based on the knowledge of rehabilitation medicine, combined with the results of motion recognition and prediction, the evaluation model of gait rehabilitation training effects needs to be established, and the method of generating personalized rehabilitation training prescriptions needs to be studied. Based on the extracted non-steady-state motion signals of the lower limbs, the time–frequency characteristics of the vital signs signals such as EMG signals can be analyzed. In order to evaluate the movement synergy of the healthy and abnormal limbs of the human body, the mechanism of human muscle synergy needs to be further studied. Combining the lower extremity musculoskeletal model with the static optimization algorithm to calculate the muscle activation degree during human walking, the evaluation method of the muscle coordination degree on the lower extremity muscle movement coordination ability of the patient's exercise training can be studied. Finally, the evaluation of gait training effects for different ages and different training stages will be realized.

The workflow of an ideal gait training system should be as follows: based on the evaluation of walking ability and the needs of gait rehabilitation training, the movement mode of gait training can be determined. The corresponding human motion intention and motion reference trajectory are obtained through the non-contact motion sensing system. After this, the desired motion trajectory of the gait training system is generated. Combined with the motion intention of the lower limbs of the human body and some simple control methods, the gait training system will flexibly assist the patient to complete the desired action. All in all, the key to the formulation of control strategies is gait evaluation and intention recognition, while obscure and sophisticated control theory is secondary. By studying the collaborative control method of the gait rehabilitation training system and the patient, based on principal component analysis, multiple regression, and neural network, the association model between gait data and clinical evaluation can be constructed, and a personalized gait training strategy with multi-layer, and cooperative closed-loop control of "human in the loop" can be designed. Based on this, carrying out research on the collaborative control of human–machine systems based on personalized rehabilitation strategies, evaluating the perception and control performance of the gait training system, and generating clinical evaluation reports on the effects of rehabilitation training have important academic significance and extensive clinical application value.

**Funding:** This work was supported in part by the National Natural Science Foundation of China under Grant 52175033, U21A20120 and U1913601; the Zhejiang Provincial Natural Science Foundation of China under Grant No. LZ20E050002; Open Fund of the State Key Laboratory of Fluid Power and Mechatronic Systems: GZKF-202101; DongGuan Innovative Research Team Program (2020607202006).

**Institutional Review Board Statement:** Not applicable.

**Informed Consent Statement:** Not applicable.

**Data Availability Statement:** Not applicable.

**Conflicts of Interest:** The authors declare no conflict of interest.

#### **References**

