Autonomous balance control of underactuated torque-controlled lower-limb exoskeletons for paraplegic users
Ander Vallinas Prieto is a PhD student in the Department Biomechatronics and Rehabilitation Technology. (Co)Promotors are prof.dr.ir. H. van der Kooij, dr. E.H.F. Asseldonk and dr.ir. A.Q.L. Keemink from the Faculty of Engineering Technology.
Balance control is key to every motion humans make and to every posture humans hold. The pull of gravity, reaction forces to our own motion, and external disturbances compromise balance and need to be compensated to prevent balance loss. Muscles work in synergy in an autonomous manner to achieve balance control while performing the desired motion, and this control is often taken for granted due to its involuntary nature. However, people with complete spinal cord injury (cSCI) lack motor function below their lesion and are unable to control their lower limbs. Thus, they cannot stand, walk, or balance using their legs. The loss of autonomy leads to muscle atrophy and other secondary health complications, which include a toll on mental health.
A viable solution to allow people with cSCI to regain their ambulatory capabilities is the use of lower-limb exoskeletons (LLE). These devices are wearable robots that apply moments at the joints of the user to drive their legs. LLEs can be categorized by the joints of the user that they drive or how the actuators that drive these joints are controlled. In terms of actuator control, we can identify torque-controlled and position-controlled LLEs. The latter are more prevalent and impose predefined motion profiles that allow users to walk. However, these devices, with a couple of notable exceptions, require the user to apply the forces necessary to maintain balance with the muscles of their arms utilizing crutches, which is quite exhausting. Furthermore, users hold the crutches with their hands, so they cannot use them to physically interact with their environment and perform daily life activities. These shortcomings limit the use and appeal of LLEs for cSCI subjects, and the inclusion of balance control in the LLE control can address them.
In contrast to position-controlled LLEs, which impose motion profiles at the joints, torque-controlled LLEs can impose joint moments. Desired moments can be prescribed through model-based computations to accelerate the joints in a desired manner while counteracting the effects of gravity, reaction forces due to planned motion, and deviations caused by disturbances in a coordinated way that prioritizes maintaining balance. This has been achieved in humanoid robots by controlling the centroidal dynamics of the robot, which is to say, by applying forces and moments at the center of mass (CoM) of the robot to steer the linear and angular momentum in a controlled manner. This approach is the so-called momentum-based control (MBC). MBC needs to be tailored to the controlled device, and the application of forces and moments at the CoM is non-trivial: the amount of joints that the LLE drives as well as the upper and lower limits of the applicable moments, whole-body posture, contact configuration and friction at the contact interfaces severely influences the magnitude of forces and moments that are produced at the CoM.
The main objective of this thesis is to develop an autonomous MBC balance controller for an LLE that can support a person with cSCI. This thesis investigates the feasibility, design, and validity of such a controller in a torque-controlled LLE.
In this thesis, we first investigated the feasibility of MBC in our setup. To that end, we developed a method that computes the set of feasible moments and forces at the CoM of a robot given the model of the robot and its limits. The used LLE is the Symbitron exoskeleton, which has four torque-controlled joints per leg (hip ab-/adduction, hip flexion, knee flexion, and ankle flexion) and a passive joint aligned with the ankle in-/eversion, which makes the LLE underactuated. The set of feasible forces and moments is dependent on the robot pose and contact configuration, so we evaluate this set for our LLE in two static standing poses: parallel double stance and single stance. For parallel double stance, the results suggest that the Symbitron can apply forces and moments in all directions except a moment around the vertical. For single stance, the Symbitron cannot produce a moment around the normal vector of the frontal plane coupled with a force in the mediolateral direction, due to the passive nature of the ankle in-/eversion joint. We can conclude that disturbances in these `non-actuated' directions cannot be instantaneously rejected but require a change in pose and contacts. Nevertheless, forces and moments computed using MBC to reject perturbations and achieve balance during double stance are feasible by the Symbitron LLE.
Next, we designed a MBC balance controller that can be deployed in the Symbitron LLE. The controller fulfills various control goals that are designed to keep balance and the contact configuration of the device: bring the CoM to a reference position, regulate the torso to a reference orientation, and minimize angular momentum and feet velocities. All control goals are fulfilled simultaneously with different priorities determined by numerical weights.
We first tested the controller performance in simulation with two levels of realism: with ideal sensors and perfect torque-source actuators and with noisy sensors and finite bandwidth actuators with their low-level control. The simulated LLE was subject to external disturbances (pushes at the torso in the mediolateral and anteroposterior directions) while standing in parallel and staggered double stances. The MBC standing balance controller was used to compute joint torques and the perturbations were successfully rejected, maintaining balance in all tested scenarios. Furthermore, the simulated LLE was commanded to follow a CoM trajectory that shifts the weight between the feet and was also able to maintain balance during this motion.
With the real robot (without a user), we first repeated the same set of experiments. The Symbitron LLE was capable of self-balancing while following a weight-shifting trajectory and rejected pushes up to 30~N both in parallel and staggered stance. Subsequently, we evaluated the disturbance rejection of the LLE for different CoM control gains to get insight into controller tuning and improve performance. The derived heuristic guideline is that, although the CoM control should not be too compliant, the controller stiffness and damping should not cause device instability and advanced balance recovery strategies should be implemented in challenging/extreme cases. The proposed balance controller is a valid approach to computing joint torques that preserve the balance of this underactuated LLE.
Last, we present the experimental validation of the balance controller implemented in the Symbitron LLE with two non-disabled and one cSCI user. We further improved state estimation and controller design to ameliorate the stability of the controlled system and account for user dynamics in the internal model. The human-exoskeleton system (HES) can withstand pushes in the order of 60~N while in parallel stance, staggered stance, and staggered stance with the front foot at a height 5~cm higher than the trailing foot. Thus, the proposed MBC balance controller is a valid approach to keep the balance of people with cSCI during stance with a torque-controlled, underactuated LLE in three different poses that appear as resting stances and during locomotion on flat ground and uneven terrain.
To finish this summary, in this thesis, we present a method to evaluate LLE design and suitability for control strategies and a controller architecture that successfully keeps the balance of a person with cSCI. Furthermore, with this control architecture, the addition of new control objectives and modification of references, gains, or weights of the existing objectives is trivial and can be done during execution. Thus, the controller is readily extendable to trigger more complex balance recovery maneuvers like reactive steps, and transition to actual locomotion in the future. The validation of this controller and the acquired insight into its inner workings are great and promising steps toward restoring the autonomy of people with cSCI.
