The functional characteristics drives the exoskeleton design process. Only knee and ankle joint are actuated, given the preservation of hip flexors. The exoskeleton can be used by subjects up to 90 kg. The exoskeleton frame is based on a bi-lateral structure hinged at the knee and ankle joints, with a single external support bar, made of Aluminium 7075. It is customizable to the user due to telescopic rods built in the distal and proximal segments. Transmission of forces between the subject and the exoskeleton is secured by four girths and velcro straps, two at the thigh and two at the shank respectively.

The knee joint mechanism is made up of four-bar bio-inspired mechanism [10,11]. This mechanism allows the joint axis follow the translational motion of the physiological knee joint axis, improving kinematic compatibility between the orthosis and the subject with respect to a single-axis joint. This mechanism was modified to incorporate the actuator, comprising a DC flat motor (Maxon EC-90) and an Harmonic-Drive transmission (ratio 100:1). Motor and transmission were selected using non-pathological gait data at slow speed (0.5 m/s [12]), from which maximum torque and power to drive the knee were extracted.

The ankle joint actuation is directed to avoid the droop foot during swing phase of walking. This function is implemented by a torsional spring, which allows a lightweight and compact design. The stiffness of the spring has been calculated from anthropometric data particularized for a 90 kg user [12], in order to avoid drop foot during the swing phase, which resulted in 1.4 Nm/°. Such stiffness allows plantar and dorsal flexion during stance phases of gait with a low interference with the movement. The exoskeleton prototype is shown in Figure 1(a).

Most of the sensory system is aimed to monitor the events and states of walking. As stated in Section 2.4, online interaction analysis makes necessary to complete the sensory system with a force sensor, which is shown in Section 2.3. An scheme of the hybrid exoskeleton’s sensory system is shown in Figure 1(b).

Normal gait can be described as a series of states and transitions among them. Main states are stance, when the foot is contacting the floor, and swing, when the foot does not contact with the floor. These two main states impose different kinematic joint trajectory and stiffness to the exoskeleton joint actuators so the controller must be fed with the information regarding foot contact with the floor. Therefore we have implemented two Force Sensing Resistors (FSR) (Interlink Electronics Inc., 38.1 mm × 38.1 mm), placed at the heel and tip of the exoskeleton’s foot insoles, which provides an ON/OFF signal to the controller to discriminate between swing and stance.

As mentioned, each gait state imposes a kinematic reference to the exoskeleton’s joint actuator, so it is necessary to monitor the actual joint angle. In our design, only knee joint is actively controlled, while ankle joint is passively controlled by a torsional spring. Monitoring of knee angle is done trough a rotational potentiometer (Vishay Espectrol). In order to assess gait biomechanics, we measure also shank and trunk orientation as well as hip joint angle. Orientation of shank segment is measured by an inertial measurement unit (TechIMU, TechnAid S.L.), which gives the 3D orientation in real time. This inertial measurement unit (IMU) uses information from a 3D accelerometer, a 3D gyroscope and a 3D magnetometer to calculate the orientation of the IMU in a world-referenced coordinate system. Another IMU is used to measure trunk orientation and, combined with the shank IMU, we calculate the 3D hip angle. Also, the trunk IMU is used to monitor trunk orientation, in order to provide information about user’s equilibrium and intention: if the trunk is tilted forward, the subject aims to keep walking, while if the trunk is extended or tilted backward, the subject aims to stop walking.

This set of sensors, in combination with the knee actuator, the ankle passive actuator and an adequate control strategy, allows to achieve an automatic walking. However, it has been described that the users of this technology does not feel comfortable with this kind of automatic behavior [13]. Therefore, volitional control of the hybrid system is accomplished by two hand-switches, implemented by a small force-sensing resistor. This implementation allows placing a hand switch on the user’s grip of crutches or walker without interfering with it. Orders implemented are: sit to stand, right leg step, left leg step, stand, and stand to sit. Those orders are combined with the information form the inertial sensor placed at the trunk to avoid dangerous situations (i.e., order of “step” when the trunk is tilted backwards)

Extracting information regarding physical interaction can be achieved via a force sensor placed on the exoskeleton structure. The interaction force is the force developed at the leg-exoskeleton interface during the movement, that is, the girths’ inner surface over the leg skin. This force, depending on the distance at the girths are placed for a specific subject, can be transformed to a torque at the knee joint level.

There is a broad offer of commercially available force sensors. However, it is difficult to find an adequate force sensor, with adequate geometry and range of measure, that can be placed at the exoskeleton structure. Moreover, integration of a force sensor will result on introduction of discontinuities on the structure. Therefore, we instrumented the structure to turn the structure into the force sensor itself.

In order to instrument the structure, we have introduced a full gauge bridge over its distal part, below the knee joint and over the first girth of the distal segment of the exoskeleton. Measurement of forces through strain gauge is based on measuring the deformation of the structure caused by the own force [14]. Therefore, the exoskeleton structure must deform adequately in order to measure the interaction torque. Of course, this deformation should be enough to provide adequate resolution, but must not be excessive to not overcome the strain limit of the aluminium.

We performed an analysis of the exoskeleton structure in order to design most adequate geometry to achieve a desirable deformation. This analysis had a two-fold objective: first, the geometry would should allow to place a full gauge bridge. Second, an adequate range of measure must be achieved, related with the forces involved during walking. An estimation of knee joint moment during non-pathological gait at low speed [12] was taken to perform the analysis. Other conditions assumed were the following:

Muscle electrical stimulation will develop 25% of maximum knee joint torque during walking.

Once the geometry was designed, a verification by finite-element analysis was conducted. A maximum strain of 0.2% was obtained (Figure 2(a)). The actual sensor is shown in Figure 2(b).

Our hybrid approach for control of walking is based on monitoring the physical interaction between the user and the exoskeleton. Although the design and implementation of the control architecture is out of the scope of this article, it is necessary a brief description. It is shown on Figure 3. The control architecture has two control loops: one performs closed loop control of knee trajectory (top loop at Figure 3), and the other control the muscle stimulator. Coordination between the two controllers (therefore the hybrid control) is performed by the state machine, which governs both controllers. This state machine (blue box) imposes kinematic pattern to the PID controller of the knee, and the stimulation pattern (defined by channel to activate and its characteristics of intensity, frequency and pulse width) to the stimulator, regarding the information given by the sensory system.

Closed loop control of muscle stimulation is difficult to perform due to the absence of appropriate signals from the muscle under stimulation that allow to perform direct closed loop control of the stimulation parameters. Instead indirect measures have been taken to close the stimulation control loop, such as joint angle due to the stimulation [15,16], the interaction torque developed due to muscle stimulation [17,18] or a combination of joint angle and interaction torque [19,20], to adjust stimulation intensity. However, none of these work has directly managed muscle fatigue due to stimulation, therefore system’s efficacy would decrease. Our approach aims to close the stimulation loop by monitoring muscle performance through user-exoskeleton physical interaction. This information will be used to modulate stimulation intensity regarding muscle performance.

We performed a preliminary experiment with a healthy subject in order to corroborate that our design is able to assess human-robot interaction for fatigue estimation. Muscle performance is assessed from physical user-exoskeleton interaction under the hypothesis that user’s joint torque developed by muscle electrical stimulation will decrease due to fatigue, as the muscle will generate less force with time while the stimulation pattern (intensity, pulse width and frequency) is kept unchanged. This reduction in muscle force must be counterbalanced by the exoskeleton actuator, applying more force to the leg, which will result in an increase of the interaction torque. Therefore, interaction torque can be used to estimate muscle fatigue due to electrical stimulation.

In order to have a more robust muscle fatigue estimate than instantaneous interaction torque, we have developed a metric that would represent the mean interaction torque produced during a time interval. To provide an adequate basis of comparison, we have focused only on the swing phase of gait because, considering also that the stance phase would introduce uncertainties into force readings as the user can support its own weight with the other leg (or crutches if it is a SCI user), which may vary among steps. Therefore, as the muscle fatigue index is applied to swing phase, the experiments performed were focused on such phase.

This mean interaction torque is calculated by equations (1) (mean interaction torque during leg flexion) and (2) (mean interaction torque during leg extension). In a first approach, we have analyzed the mean interaction torque during flexion and extension separately given that the muscles involved in each movement are different. (1) T flex ¯ = ∫ flex T int ⋅ dt t flex (2) T ext ¯ = ∫ ext T int ⋅ dt t ext

In those expressions, T flex ¯ and T ext ¯ are the mean interaction force, calculated by integrating the interaction force measured by the sensor (T_int), during both flexion and extension intervals, averaged by the interval time (t_flex and t_ext, respectively). This subject did not practice sports at least for 24 h prior to the experiment. The experiments lasted 10 min, with the leg free to swing. User was requested to not to move the leg and try to relax during the experiment, in order to avoid any voluntary movement. The kinematic pattern fed to the controller was a series of flexion-extension trajectories from a non-pathologic gait kinematics database, that is, only the swing phase of gait. This kinematic pattern was slightly accelerated to shorten the experiment duration.

While the exoskeleton controls the joint trajectory, a synchronized stimulator (Unafet8) delivered stimulation pulses to vastus lateralis muscle during extension and biceps femoris muscle during flexion. Stimulation pulses was monophasic compensated, 300 ms pulse width and 40 Hz frequency. Stimulation intensity was set regarding user comfort perception, but high enough to achieve a functional knee joint movement. The intensity achieved was 46 mA for flexors muscles and 28 mA for extensors muscles. The delivered stimulation pulses were synchronized with the kinematic pattern. Figure 4 shows the experiment setup.