The somatosensory function includes the senses of touch, vibration, pressure, proprioception, pain, and temperature. Impairments in this function negatively affect the ability to perceive, distinguish, and recognize the senses in the body (Aries et al., 2021). Consequently, disorientation accompanied by abnormal movements, balance disorders, muscle weakness, and inability to maintain postural control (stabilize the body against gravity and perturbation) may occur (Kim and Jang, 2021). For instance, it is reported that post-stroke individuals experience high rates of somatosensory impairment, ranging between 65% and 85% (Costantino et al., 2017). A research with self-questionnaire demonstrated that individuals with somatosensory and motor impairments suffer from lower walking capacities and lower levels of independence in ADL (Patel et al., 2000; Tyson et al., 2013; Gorst et al., 2019). Since restoring walking ability is a primary objective for many stroke patients, establishing the best treatments for balance, gait, and mobility were identified as one of the top ten stroke research priorities (Sánchez-Blanco et al., 1999). In a recent survey conducted with 145 stroke individuals, 43% of individuals reported decreased sensation in their feet; sensory impairment was indicated to be the second most common foot problem after the loss of strength (Bowen et al., 2016; Gorst et al., 2016). Limitations in walking, high fall rate, and impairments in foot-ground contact and sense of foot position sense and hence decrease the outdoor activities in the community.

Robot-assisted rehabilitation is continuously gaining prominence as it provides more efficient training and objective evaluation than conventional rehabilitation procedures (Saglia et al., 2009, 2010; Shakti et al., 2018). Furthermore, conventional methods require at least three physiotherapists to support the lower extremity and trunk of the patient manually. Another limitation is that the effectiveness of the rehabilitation during these practices depends on the personal knowledge and experience of the therapist. It is reported that the demand for physiotherapists is increasing continuously to match the number of patients due to the increase in the aging population worldwide. The use of robotic devices to address these challenges is encouraged to shift the adoption of rehabilitation clinics from conventional methods to robotic-assisted rehabilitation. Hence, high-quality therapy sessions can be achieved at a relatively low cost and with significantly less effort (Díaz et al., 2011; Krebs and Volpe, 2013; Yurkewich et al., 2015; Kalita et al., 2021).

Most neurological patients have a reduced range of motion (ROM) in their ankle as well as muscle strength in their bodies. In the literature, muscle-strengthening for ankle rehabilitation has been considered with three main phases: ROM, strength, and proprioceptive training, respectively, and they require different control methods. At the beginning of the rehabilitation, the patients have limited ankle mobility; thus, passive ROM exercises are required (the patient is passive and the device is fully active) so that ankle ROM can be recovered via full assistance of the device. Then, active ROM exercise is utilized by reducing the level of assistance from device, and the patient should put effort to initiate the motion against the device. The second phase is about improving the ankle stability; as the patient's response improves, the resistance level of these strengthening activities increases. In most cases, they are unable to exert enough force to complete the exercises, so the robot should assist them. In the final phase, proprioceptive training should be performed to improve the balance control (Bernhardt et al., 2005; Valles et al., 2017; Teramae et al., 2018). However, due to their limited design configuration, that is, low weight-bearing capacity or small end effector area, most rehabilitation platforms cannot be used in the balance rehabilitation; the detailed description is given in Section 2.3. The control strategies in the stated phases are also illustrated in Table 1.

Even though robotic-assisted rehabilitation has a promising effect on the treatment itself, especially with the capability of being repetitive and task-specific, implementation of these potentials at the desired level is still challenging. There are numerous methods for control of these robots, yet control-related challenges are to be addressed (Marchal-Crespo and Reinkensmeyer, 2009; Dzahir and Yamamoto, 2014; Li et al., 2017; Jin et al., 2018; Dong et al., 2021). Position control is one of the most used methods due to its simplicity. In this approach, the robot tries to follow a predefined motion trajectory (Mohebbi, 2020). Furthermore, the position control is needed to conduct the movements in the first phase along a specific trajectory at a constant speed to improve the ROM capacity of the patient (Girone et al., 1999; Saglia et al., 2013; Zhang et al., 2016; Ayas and Altas, 2017). However, this method cannot be used in the other phases since when the patient applies more than the expected force/torque during rehabilitation, the position method cannot compensate for this disruptive effect, which leads to excessive position tracking errors. Despite this, it should be noted that most of the existing interaction controllers continue to use a position control scheme as an inner control loop, with a corresponding force/torque or motion outer loop applied to complete the interaction controller (Lu et al., 2016; Song et al., 2019).

To eliminate the problem with position control, impedance control, which provides the desired dynamic interaction between the robot and its environment, is proposed. The physical interaction is expressed as the dynamic relationship between the motion variables of the manipulators and the contact forces that need to be kept within a predetermined safe or acceptable position trajectory while the robot follows the desired motion trajectory (Song et al., 2019). Admittance control uses force as an input and displacement as an output, while the robot follows predetermined force. These methods are used in the strength phase to assist/resist the patients (Saglia et al., 2013; Ibarra and Siqueira, 2015; Sun et al., 2015; Jamwal et al., 2016). However, these methods require force/torque sensor to measure the force exerted by patients, which adds extra complexity, that is, sensor dynamics, cost, and computational load. Moreover, the estimation of impedance parameters peculiar to the individual is another challenge (Codourey, 2016).

The dynamic model of robots is usually composed of non-linear functions of the state variables (joints' positions and velocities), especially in parallel structures. This characteristic of the dynamic model makes the closed-loop control system non-linear and difficult to solve. The computed torque control (CTC) method requires a good knowledge of the robot dynamics since the dynamic model of the manipulator is used in the loop. Even though CTC can compensate for non-linearity since dynamic equations are solved in real time, it creates a computational burden (Tsoi et al., 2009; Asgari and Ardestani, 2015; Codourey, 2016).

The above-mentioned issues are addressed by using the parameter estimation control method. With this method, the simplification of complex mathematical dynamic equations and modeling of unmodeled noise signals are done and a new physical model with functional properties can be obtained with less computation time and acceptable control performance. Furthermore, modeling can be done online or offline. In the offline method, if the estimated model is not highly accurate, it cannot be able to correctly distinguish between responses caused by known and unknown input signals. In the online method, parameter estimation is done simultaneously within the process, causing a delay in the system response (Wolbrecht et al., 2008; Gao et al., 2014; Song et al., 2019). The fuzzy logic method is also used in the control of rehabilitation robots. Furthermore, it provides good performance for the control of the non-linear system. Fuzzy logic is similar to human thought systematic than traditional logic systems. Basically, it tries to capture the approximate, imprecise nature of the real world. It has three steps: fuzzification, linguistic rules, and defuzzification. In the fuzzification step, input signals are converted into fuzzy sets with some degree of membership (range from 0 to 1). The main part of fuzzy logic control is controlling the robot using a set of linguistic control rules associated with binary concepts such as fuzzy inference and computational inference rules. In the defuzzification, fuzzy truth values are converted into output decision values. However, like parameter estimation, it is not robust against unexpected situations during real-time control since its rules are determined previously (Karasakal et al., 2005; Lamamra et al., 2020; Sharma and Obaid, 2020).

In the above-mentioned methods, the patient passively follows the movements of the robot, which follows the previously specified position or force/torque references (Chiaverini and Sciavicco, 1993; Patarinski and Botev, 1993). Recent studies in robotic therapy show that continuous passive training therapy does not significantly improve motor function. Active participation of patients is considered to be a major factor contributing to the neural plasticity and motor recovery (Keller et al., 2013; Teramae et al., 2018). One of the most commonly adopted assistance strategies, the “Assist-as-Needed (AAN)” paradigm, offers the mode of necessary active assistance to stimulate neuroplasticity. The basic principle in AAN is to provide physical assistance only when needed by the patient. If a patient performs a task flawlessly, the robotic assistance is withdrawn. However, if the patient has difficulty or cannot complete the task, the robot provides as much support as the patient needs to perform the task (Pehlivan et al., 2017). In other words, AAN is a strategy of regulating auxiliary forces/torques or task difficulty according to patients' disability level or performance in training tasks. There is strong evidence that active participation induces neural plasticity, and therefore controllers should intervene minimally to promote participation and recovery. In addition, upper extremity rehabilitation using the AAN paradigm is shown to be the most promising technique for promoting recovery (Luo et al., 2019).

New technological devices and methods are being developed to increase active patient participation by integrating multi-sensory information and allowing AAN paradigm. However, the implementation of AAN paradigm is still challenging, since determining the level of assistance according to the patient progress is not to be addressed sufficiently.

Considering the design and development of robotic devices for lower extremity rehabilitation, there are mainly three types of system, that is, wearable exoskeleton system, ankle platform, and balance rehabilitation platform are proposed (Deng et al., 2018; Ersoy and Hocaoglu, 2023). Several examples of exoskeleton devices aim to increase the capacity of the lower limbs or reducing user effort (Pratt et al., 2004; Ferris et al., 2006; Dollar and Herr, 2008; Mooney et al., 2014). Systems such as AKROD (Weinberg et al., 2007), BioMot project (Bacek et al., 2017), KNEXO (Beyl et al., 2009), Lokomat (Jezernik et al., 2003), LOPES (Van Der Kooij et al., 2006), MIRAD project (Mir, 2022), and REX (Rex, 2022) are used to support individuals with muscle weakness in ADL. However, since the mechanisms underlying human movements and how the designed devices should interact with humans are not fully understood, there is no device that can effectively improve the user's performance. The mentioned systems have difficulties in use because they have a rigid, bulky structure, and uncomfortable interfaces, restrict biological joints, and are misaligned with natural joints. In addition, if the exoskeleton does not have enough degree of freedom (DoF) to work in harmony with human joints, it exerts a residual force on the human limb due to axial misalignment, and this may cause long-term injuries, as well as discomfort (Schiele, 2008, 2009).

Platform-based robots are grounded and have movable end effector as a rehabilitation platform with one or more DoF. These types of systems employed for the ankle joint focus only on improving the ROM of the joint rather than improving the balance of the patient (Díaz et al., 2011; Morris et al., 2011). Most of the platforms for the ankle joint are in parallel structure, which provides sufficiently high torque for plantar flexion/dorsiflexion, inversion/eversion, and adduction/abduction movements of the ankle (Chablat and Wenger, 1998; Rastegarpanah et al., 2016). There are also systems designed to be serially connected to each other with motor-operated joints (Saglia et al., 2019). Although serial manipulators are easier to model, using a parallel manipulator in such an application provides advantages in terms of achieving high load-carrying capability, better dynamic performance, and precise positioning. Rutgers ankle is a Stewart platform-type haptic interface that provides 6 DoF resistance forces to the patient's foot in response to VR-based exercises (Girone et al., 2001). Various clinical studies are conducted with this device showing improvement in patient strength and endurance measurements (Girone et al., 2000; Deutsch et al., 2001; Cioi et al., 2011; Deuschl et al., 2020). In RePAiR, a 1-DoF robotic platform allows dorsiflexion and plantar flexion movements to patients after stroke. It is demonstrated that the device provides benefits in increasing the muscle strength of the patients, improving the motor control, sensory-motor coordination of the patients, and, accordingly, the walking patterns (Gon calves et al., 2014). Only a few of the manipulators developed for ankle rehabilitation are commercialized (Saglia et al., 2013; Bre, 2022; Opt, 2022). These platforms are widely used to strengthen ankle joint movements and improve ankle proprioception. The end effector of the above-mentioned manipulators are only one foot large and have a low weight-bearing capacity; therefore, they cannot be used for balance rehabilitation after ankle treatment has been completed. In addition, the end effector of the proposed systems are not endowed with a sensor; therefore, pressure change measurement on the sole and sensory input under the foot cannot be performed during the ROM rehabilitation.

Posturography, measurement of CoM and balance variables, is tested through static or dynamic techniques for balance evaluation and rehabilitation (Prosperini and Pozzilli, 2013; Park and Lee, 2014). Training with static balance platforms is said to be helpful in controlling pressure distribution in patients over time. Static posturography is usually done by using Wii fit (Wii, 2021) board or commercially available force platforms (AMT, 2022; Ber, 2022; HUR, 2022; Kis, 2022). Studies evaluated the Wii exercise experience of the patients and their physiotherapists, are stated that Wii exercise is an amusing and challenging way to improve balance impairment since VR provides continuous feedback (Prosperini and Pozzilli, 2013; Plow and Finlayson, 2014). Another study was conducted for 6 week with commercially available force platform and virtual reality and the result shows that the CoM point deviation decreased. However, when compared to static situations, rehabilitation under dynamic conditions contributes more to the improvement of balance disorders and motor skills (Prosperini and Pozzilli, 2013).

Dynamic platforms require instantaneous dynamic movements, forcing patients to adjust their balance during perturbation (Prosperini and Pozzilli, 2013). A study shows that dynamic strength platforms are more helpful in restoring postural stability than conventional therapy (Saglia et al., 2019). Wooden balance platforms are one of the simplest examples, but they do not provide any quantitative measurement (Bal, 2021). The Bobo Balance platform provides force measurement on top of wooden platforms. However, it is difficult to use for people with severe loss of balance since there is no external support environment or mechanisms for patient while standing (Bobo, 2021). gePRO (Geapro, 2021), BackinAction (Ria, 2022), Balanceback (Bal, 2022), and Proprio (Pro, 2022) platforms have 2-DoF (roll and pitch) for balance rehabilitation. These systems deliberately put patients in an unbalanced state while patient following the VR game, thus assessing their balance status based on CoM position. During the dynamic rehabilitation process, the angle of the platform can be controlled, and patients are requested to maintain their CoM and posture. Even though ROM can be covered fully by gradually changing the angular position, these systems cannot provide AAN paradigm (even if a patient performance improves, the system's level of assistance remains unchanged) (Marchal-Crespo and Reinkensmeyer, 2009; Kang et al., 2015). Furthermore, weight transfer in balance training cannot be provided during PBT due to the lack of sensory input to the foot with these devices. Additionally, the VR environment is the only sensory input proposed with these devices; however, it has been stated in the literature that multi-sensory input is more powerful for mimicking and enhancing ADL (Sigrist et al., 2013).

Investigation of APAs and CPAs, which is necessary for proactive and reactive balance control, is claimed to disclose essential aspects of postural control and history of falls. The studies show severe APA and CPA deficiencies after evaluating the patient's condition and muscular activity; however, retraining APAs and CPAs is possible through PBT programs. Despite the benefits of these studies, to our best of knowledge, there is no study in the literature that assesses reaction time, CoM, CoP, and the ability to control APAs and CPAs on ankle, balance, and stepping rehabilitation simultaneously. On top of that, the effects of multi-sensory inputs and cognitive control strategies suitable for the use of patients with different severity levels have not been evaluated and utilized.

The first phase is ankle-foot/preparation rehabilitation to improve lower extremity muscles, while the patient is sitting due to their ankle instability and low muscle activation. It aims to prepare the sensorimotor system for motor function, which is essential when there is minimal or no voluntary motor/muscle function in ankle-foot. The selected robotic platform should be kept in ankle ROM, shown in Figure 3A with 0–20^o dorsiflexion, 0–50^o plantarflexion, 0–10^o adduction, 0–5^o abduction, 0–20^o eversion, and 0–35^o inversion limits (Hasan and Dhingra, 2020). The system needs to ensure that each patient can perform exercises at their specific ankle ROM limits. Upper limits must be determined according to the impairment levels of the patient, and it should be gradually increased based on their performance. Such limits can be identified with the use of a robotic platform with a predetermined force/torque based on an individual's parameters, that is, foot mass, size, and inertial parameters. This torque should have the maximum magnitude to be able to move the individual's foot. Patient can exert reaction force/torque due to spasticity (involuntary muscle contraction) and RoM limits during robotic platform induced motion. Thus, once platform cannot move the foot any further, this angle is considered as patient's spasticity level. Up until this time rehabilitation is performed with patient passive mode; however, as patient progresses in the therapy, his/her contribution to the motion would also be requested at this phase (patient active mode). For instance, moving the ankle to 5^o of dorsiflexion with patient active contribution while the platform supports and this support can be regulated according to his/her progress.

Accordingly, in a study conducted with 459 people after a stroke, it is observed that 45% of the patients had a decrease in lower extremity sensation (Sarah et al., 2013). Therefore, a force and haptic feedback sensor should be placed in at least four regions, namely, toes, metatarsals, middle foot, and heel, as they are more intense regions (Perttunen, 2002) (see Figure 3B for feet placement and see Figure 3C for specified regions). With the force sensor, the pressure applied by the patient to the each region on the sole can be measured and if the pressure distribution exceeds the able-bodied pressure distribution value in one of these regions, he/she can be stimulated that region via haptic feedback based on the individually identified sensing threshold. Accordingly, the patient can be aware of their amount of weight-bearing and sole pressure; thus, they can be trained to transfer their weight correctly for fall prevention. With this aim, a personalized preparation therapy program is proposed as an essential stage for I-BaR, which provides multi-modal feedback/feedforward signal and determines the somatosensory level of patient for the therapist to plan and decide on the parameters of the next phase.

The second phase is balance rehabilitation to improve postural adjustments while the patient is standing on the platform. Although the patient pass to this phase after the preparation, they may have difficulty in balancing their posture without physical assistance. Therefore, a haptic bar should be installed on the selected rehabilitation system (see Figure 3D). This haptic bar should at least include two sensors, force and haptic sensors, to quantitatively measure applied force and give haptic feedback accordingly. In the case of a patient with higher balance loss, the phase should be executed with full assistance of haptic bars. Throughout the rehabilitation as the patient's postural control improves, an upper extremity support limit value should be established for each patient. When the patient goes over the limit, haptic feedback is applied to warn them to reduce the upper extremity support and help them to control the posture by using the lower extremity rather than relying on the arms.

As the core part of balance rehabilitation, this phase should be utilized to increase sensory and motor integration during task-specific activities. Patient-specific information, such as ROM and proprioception, obtained from the preparation phase determine the required parameters of the therapy at this phase. This rehabilitation can also increase somatosensation during task-specific activity. Within the scope of these interventions, the patient's balance in the mediolateral and anteroposterior directions can be disrupted by giving sudden and fast perturbations with rehabilitation robot. In PBT, to train balance reactions, perturbations in different directions and amplitudes are applied, and various scenarios are used, including simultaneous cognitive and motor activities. The objective of this is to decrease the reaction time and deviation in CoM and also to improve pressure distribution as well as APAs and CPAs (Mansfield et al., 2015; Jagdhane et al., 2016). As in the preparation phase, the AAN control paradigm should be implemented on the Robotic platform system during PBT so that the patient can engage with the dynamics of the platform independently and actively within the limits of their capacity.

The third phase is stepping rehabilitation to improve walking parameters, that is, the stance and swing phase's temporal and spatial parameters. While step speed is increasing, improving the stepping accuracy is another target in this phase. It should include assessment and training of stepping to various target distances with robotic help. Step-taking activities are performed to assess the effectiveness of the previous two phases by analyzing the relationship between muscle activity, postural control, movement speed, and accuracy as well as to train walking parameter with Fitts' law. Particularly, at stepping long distances, the patient realize the movement faster with their own compensation strategies which reduces the quality of motion, whereas at stepping short distances their motion is slower, controlled and precise. Current step-taking assessments are based on manual change of the target distance and reaction time measurements (detected by a switch button according to the time between target point and initial position force detection) (Aloraini et al., 2019, 2020; Yamada and Shinya, 2021). Since force measurements and haptic sensors are not implemented at these target points, a kinetic evaluation and improvement in sensory input under the foot cannot be provided. Therefore, in the stepping phase, the distance of the target points should be automatically adjustable within the system to provide perturbation/moving target in x-y plane and these target points should include force and haptic sensors to calculate GRF and increase sensation under the sole, respectively (see Figure 3E).

This study proposes a completely personalized rehabilitation system for patients with different severity levels or disorders. It integrates three main phases of the I-BaR framework, that is, ankle-foot, balance, and stepping rehabilitation. The originality of the study includes the following points: