Parameter Personalization of Human Arm Dynamics Model for Optimal Assistance of Assistive Robotic Arm Extender (ARAE)

Lee Hyunwoo

ARAE Original Workflow

Problem

• Overestimation of parameter from nominal model
• Overestimation of required torque
• Adversarial effect from ARAE assistance
• Instability of robot

Problem

• Overestimation of parameter from nominal model
• Overestimation of required torque
• Adversarial effect from ARAE assistance
• Instability of robot
• Inaccurate shoulder position

Project Scope

Problem Formulation

Sagittal plane model

ARAE

Estimated motor torques

Robot joint angles

Nominal human anthropometric data

Problem Formulation

Corrected Sagittal plane model

ARAE

Improved

motor torques

Robot joint angles

Personalized human anthropometric data

Problem Formulation

Corrected sagittal plane model

ARAE

Robot joint angles

Personalized human anthropometric data

Position controller

Robot joint angles

Linear least square minimizer

Theoretical motor torque expressions

Actual motor torques

MuJoCo Simulation

Purpose

• Validate corrected sagittal plane model
• Actual shoulder position from simulation VS Shoulder position calculated from old/corrected model
• Validate personalization calculation
• Torque from simulation PID controller VS Torque calculated from nominal/personalized model

Simulation Results: �Comparison of shoulder position in corrected model and original model

Shoulder position calculated from the corrected model is significantly more accurate than old model when compared with actual shoulder position

Simulation Results: Personalized parameter estimation

• Deviation between nominal and personalized parameters
• Torque calculated using personalized model is better than torque calculated using nominal model when compared to torque from PID controller

Human subject experiment

1. Personalized parameter estimation

ARAE position controller brings human arm to 6 different positions

• 6 positions to cover average daily life activities region

Torque required at each position is used for personalization calculation

2. Performance comparison between personalized model and nominal model using sEMG

3 collection modes

• No robot (baseline)
• Nominal model
• Personalized model

4 activities in each collection modes

• Forward reach (FR)
• Lateral reach (LR)
• Drink
• Scoop

Evaluation Metrics

• Features extracted from sEMG data
• Mean Average Value (MAV): Muscle activation
• Median Frequency (MDF): Muscle fatigue
• Compare change in features
• Nominal VS Baseline
• Personalized VS Baseline

sEMG Feature Analysis

Beneficiary Effect

• Decreased MAV: Reduced muscle activation
• Increased MDF: Less muscle fatigue

Adversarial Effect

• Increased MAV: Increased muscle activation
• Decreased MDF: More muscle fatigue

Human Testing: sEMG Data Collection

• Muscles Used
1. Biceps Brachii (Long head)
2. Anterior Deltoid
3. Pectoralis Major
4. Triceps Brachii (Long head)
5. Posterior Deltoid
6. Latissimus Dorsi
• 10 subjects recruited
• Electrode Placement according to SENIAM guideline

Human Testing Results: Parameter Personalization

• Significant difference between nominal and personalized parameters
• Torque calculated from personalized model is more accurate than torque calculated from nominal model as compared to torque from position controller
• Torque of motor 1 is close to zero after personalization (more stable)

Human Testing Results: Performance Comparison

• Beneficiary effect on muscles used for shoulder and elbow extension
• Adversarial effect on muscles used for shoulder and elbow flexion
• Minimization of adversarial effect while keeping the beneficiary effect

Plans

• Further refine sEMG data analysis
• FYP report writing
• Publication preparation
• Final presentation preparation

Agenda

Thank You

Lee Hyunwoo

Human Testing: Experimental Setup

• ARAE PD position controller
• ARAE, chair, and table marker
• Table marking for activities during sEMG data collection