Using Intent Estimation and Decision Theory to Support Lifting Motions with a Quasi-Passive Hip Exoskeleton

Abstract

This paper compares three controllers for quasi-passive exoskeletons. The Utility Maximizing Controller (UMC) uses intent estimation to recognize user motions and decision theory to activate the support mechanism. The intent estimation algorithm requires demonstrations for each motion to be recognized. Depending on what motion is recognized, different control signals are sent to the exoskeleton. The Extended UMC (E-UMC) adds a calibration step and a velocity module to trigger the UMC. As a benchmark, and to compare the behavior of the controllers irrespective of the hardware, a Passive Exoskeleton Controller (PEC) is developed as well. The controllers were implemented on a hip exoskeleton and evaluated in a user study consisting of two phases. First, demonstrations of three motions were recorded: squat, stoop left and stoop right. Afterwards, the controllers were evaluated. The E-UMC combines benefits from the UMC and the PEC, confirming the need for the two extensions. The E-UMC discriminates between the three motions and does not generate false positives for previously unseen motions such as stair walking. The proposed methods can also be applied to support other motions.

Figures (12)

• Figure 1: Top: A side and back view of the quasi-passive exoskeleton. Bottom: The three lifting motions to be recognized by the exoskeleton: Asymmetric stoop right, squat and asymmetric stoop left.
• Figure 2: Cross-section view of the left hip module presenting the operation principle of the actuator. When disengaged, the cam link (gray) follows the motion of the output link without compressing the springs. When engaged, the cam link is grounded to the input link. Any subsequent hip flexion will increase $\alpha$ which results in torque being generated by the springs.
• Figure 3: Side view of the left side locking mechanism used to engage the springs. The ratchet is connected to the cam link (in dark gray) via a key. The solenoid is attached to the input link. When activated, the solenoid pulls on the pawl which locks the ratchet. The mechanism is designed to be self-locking such that current only needs to be applied during the locking phase.
• Figure 4: Assistive torque provided by one hip module as function of the deflection $\alpha$ (angle between the output link and the cam link) for several levels of spring precompression $P$. The spring precompression is reported as percentage of the maximal compression of the spring which is 22mm.
• Figure 5: Overview of the (Extended) Utility Maximizing Controller. The subscript $k$ refers to the discrete time index. Raw encoder data is sent from the exoskeleton to a low-level controller where hip flexion angles $\boldsymbol{q} _k$ are calculated. In the intent estimation step, the controller will first perform onset detection and subsequently estimate model probabilities. In the decision step, the controller decides to take the action that maximizes utility to a user. The green blocks represent model data that is loaded at startup and used at runtime by the controller. The blocks in dashed lines (the calibration and velocity modules) are the extensions added to the Utility Maximizing Controller to arrive at the Extended Utility Maximizing Controller.