Development of a Bayesian Optimization Controller to Minimize the Metabolic Cost of Walking for

Human-in-the-Loop Optimization of Ankle-Foot Prosthesis (AFP) Stiffness

Tin-Chun Wen Xingyuan Zhou Michael Jacobson Seongmi Song Myunghee Kim
Mechanical Electrical and Mechanical Mechanical Mechanical
Engineering, UIC Computer Engineering, Engineering, UIC Engineering, UIC Engineering, UIC*
Chicago, USA UIC Chicago, USA Chicago, USA Chicago, USA
[email protected] Chicago, USA [email protected] [email protected] [email protected]
[email protected]

Abstract— Evidence suggests that the metabolic cost associated 2019-0087). The experiment started from control-off mode
with the locomotive activity of walking is dependent upon ankle (holding the motor), and HIL Bayesian optimization. After
stiffness. This stiffness can be a control parameter in an ankle-foot the HIL Bayesian optimization session, we provided the
prosthesis. Considering a unique individual with below-knee worst and best conditions to verify the HIL Bayesian
amputation and robotic ankle-foot prosthesis interaction, optimization results. We estimated the cost of walking by
individually tuned stiffness in the device may improve assistance measuring respiratory rate using the Cosmed K5. We
benefits. The individualization can be conducted through human-in- additionally collected foot pressure data.
the-loop (HIL) Bayesian optimization (BO). Here, we conducted a
pilot study of the ankle-foot prosthesis stiffness optimization III. RESULT AND DISCUSSION
through the HIL BO to minimize the cost of walking, shown by The HIL optimization identified a stiffness parameter 1.5,
metabolic cost. as an optimal parameter to minimize the cost of walking. The
Keywords— Ankle-Foot Prosthesis, Bayesian Optimization,
best control parameter (stiffness = 1.5) reduced metabolic cost
Human-in-the-loop optimization by 12.6% while the worst control parameter (stiffness = 0.54)
reduced the cost by 2.3%, compared to the control-off
I. INTRODUCTION
condition, respectively.
Ankle-foot-prosthesis can aid individuals with below- We also looked into the foot pressure as a potential cost
knee amputation (BKA) in walking [1]. Evidence suggests that function[5]. The mean peak pressure was 6960 KPa, 4601 Kpa,
a different individual with BKA might have a different optimal and 5054 Kpa for the control-off, worst, and best condition,
prosthesis condition [2]. Here, we tested the feasibility of the respectively. The result suggests that the pressure could serve as
individualization method through a pilot study with an a potential cost function, but more research is certainly needed.
individual with simulated BKA. We used HIL Bayesian (a) (b)
optimization to optimize a control parameter, stiffness, of an
ankle-foot prosthesis, using conventional cost function -
metabolic cost [3].

II. METHODS
A. Hardware - The AFP system consists of lightweight
ankle-foot prosthesis end-effector, improved upon
previous design [1], a Humotech Caplex Actuator Unit,
I/O and Speedgoat Performance Real-Time Target
machine.
B. Controller - The controller is composed of three layers:
low-level, mid-level, and high-level controller. The low-
level controller controls ankle-angle position, motor
position, motor velocity, and torque. The mid-level (Figure 1 (a). System setup (Pilot study with simulated
controller is a walking controller based upon gait amputation). Figure 1 (b). Optimization performance ran by
mechanics. We divided gait into two phases: swing and Bayesian optimization.)
stance phases. The stance phase is further divided into
dorsiflexion and plantarflexion. During the swing phase, IV. FUTURE WORK
we conducted ankle-angle position control. During the We will research the correlation between the cost of
stance phase, we did torque control based on the ankle- walking and the foot pressure and will study the feasibility of
angle. The overall structure of the system is shown the foot pressure as a cost function. We would like to discuss
(Fig1(a)). The high-level controller conducts the Bayesian the optimization results during Dynamic Walking 2020.
Stiffness parameter optimization using a cost function of
the metabolic cost [4]. [1] Steven H. Collins,1,2,∗, Myunghee Kim,1, Tianjian Chen1, and
Tianyao Chen3(May 26-30, 2015) An Ankle-Foot Prosthesis Emulator
C. HIL Bayesian optimization - In Fig1(b), we randomly with Control of Plantarflexion and Inversion-Eversion Torque
choose numbers from 5 to 15 to explore stiffness [2] Nicholas P. Fey, Glenn K. Klute, Richard R. Neptune1(NOV 2012)
parameters, and the stiffness value in the real controller Optimization of Prosthetic Foot Stiffness to Reduce Metabolic Cost and
should multiply 0.1. A red star indicates the previous Intact Knee Loading During Below-Knee Amputee Walking: A
Theoretical Study
parameter and a blue circle indicates the predicted
[3] RL Waters, J Perry, D Antonelli and H Hislop (1976) Energy cost of
parameter which has the biggest EI value. walking of amputees: the influence of level of amputation.
D. Experimental protocol - To test the feasibility of the [4] Jonas Mockus (2012). Bayesian approach to global optimization: theory
individualization, we conducted a pilot study. One healthy and applications. Kluwer Academic.
male subject participated in the experiment (Age: 27, [5] Sebastian I. Wolf, Merkur Alimusaj, Laetitia Fradet, Johannes Siegel,
height: 177.8cm, Weight: 84kg, IRB protocol number – Frank Braatz (Dec 2009) Pressure characteristics at the stump/socket
interface in transtibial amputees using an adaptive prosthetic foot