Reflex Control Overview

This document describes the neuromuscular reflex control system implemented in MyoAssist, based on the neural circuitry proposed by Song and Geyer (2015) for human locomotion.

Implementation Notes

This document describes the MyoAssist implementation of the Song & Geyer (2015) framework, with the following modifications:

Extended muscle set: Added FDL/EDL muscles for metatarsophalangeal (MTP) joint control
3D control: Extended to frontal plane with HAB/HAD hip muscles
Modified delays: Uses 3ms short delays (vs. 2.5ms in original paper)
Parameter scaling: Complex parameter transformations with offsets and scaling factors
Pre-stimulation: Adds 0.01 baseline activation to all muscles

Overview

The reflex control system generates diverse walking behaviors through a hierarchical structure of spinal reflex modules that emphasize sensory feedback rather than central pattern generators (CPGs). The system operates in both 2D (sagittal plane) and 3D (including frontal plane) control modes.

Neural Circuitry Architecture

The control system is organized into 10 distinct reflex modules (M1-M10), each responsible for specific aspects of locomotion. These modules use sensory feedback integration.

Stance Phase Modules (M1-M5)

M1 - Compliant Stance Leg Behavior

Uses positive force feedback from leg extensors: GLU, VAS, SOL, FDL
Provides compliant stance support and propulsion
Load-dependent modulation during stance-to-swing transitions using contralateral load factor

M2 - Knee Hyperextension Prevention

HAM: Positive force feedback for biarticular knee flexion
VAS: Negative feedback based on knee angle error (inhibited when knee > knee_off_st)
BFSH: Positive feedback based on knee angle error (activated when knee > knee_off_st)
GAS: Positive force feedback throughout stance
Load-modulated with same factor as M1 during transitions

M3 - Trunk Balance Control

HFL: Negative feedback for trunk pitch error (activated when trunk leans backward)
GLU: Positive feedback for trunk pitch error (activated when trunk leans forward)
HAM: Co-stimulation proportional to GLU activation (co-contraction mechanism)
Load-modulated by ipsilateral stance force to prevent slipping

M4 - Contralateral Swing Compensation

Activated only during contralateral swing phase
HFL: Receives input from contralateral GLU and HAM activities
GLU: Receives input from contralateral HFL and RF activities
HAM: Co-stimulation based on ipsilateral GLU activation from M4
Uses delayed afferent copies of contralateral muscle activities

M5 - Ankle and Toe Control

TA: Positive feedback based on ankle angle error, with reciprocal inhibition by SOL force during stance
EDL: Positive feedback based on MTP angle error, with reciprocal inhibition by FDL force during stance

Swing Phase Modules (M6-M10)

M6 - Swing Leg Placement Control

Uses body-frame alpha angle: α = φ_hip - 0.5 × φ_knee
HFL: Positive feedback for alpha angle error and velocity
GLU: Negative feedback for alpha angle error and velocity (antagonistic action)
Activated during swing phase and stance-to-swing transitions with contralateral load modulation

M7 - Early Swing Knee Flexion

BFSH: Negative feedback based on leg angular velocity (dalpha)
Active during swing phase 1 and stance-to-swing transitions
Ensures initial knee flexion for ground clearance

M8 - Mid-Swing Knee Control

RF: Negative knee velocity feedback (damping knee flexion)
BFSH: Combined positive knee velocity feedback AND alpha angle error feedback
Active during swing phase 2
Provides knee damping and position-dependent modulation

M9 - Late Swing Deceleration

HAM: Negative alpha angle error relative to extended target (alpha_tgt + alpha_delta)
BFSH: Threshold-based co-activation with HAM
GAS: Threshold-based co-activation with HAM
Active during swing phase 3

M10 - Final Leg Positioning

HFL: Positive hip angle error (hip flexion when behind target)
GLU: Negative hip angle error (hip extension when ahead of target)
VAS: Negative knee angle error (knee extension when too flexed)
Active during swing phase 4 when leg angular velocity reverses

3D Extensions (Frontal Plane)

M1 - HAB Force Feedback

HAB: Positive force feedback during stance with load modulation
Same activation pattern as sagittal plane M1 muscles

M3 - Frontal Plane Balance

Uses leg-specific sign convention for left/right asymmetry
HAB: Positive feedback for frontal trunk error
HAD: Negative feedback for frontal trunk error (antagonistic)

M4 - Frontal Plane Compensation

HAB: Contralateral HAB co-activation during swing
HAD: Contralateral HAD co-activation during swing

M6 - Frontal Plane Swing Control

Uses frontal alpha angle from hip adduction/abduction
HAB: Positive feedback for frontal alpha error
HAD: Negative feedback for frontal alpha error

Phase Switching Logic

The control also implements switching mechanisms (Sw1-Sw4):

Sw1: Knee flexion reaches threshold (knee_sw_tgt)
Sw2: Leg angle reaches α_target + α_delta
Sw3: Leg angle reaches α_target
Sw4: Leg angular velocity reverses (dα ≥ 0)

3D Control Extensions

For 3D control, additional modules handle frontal plane motion:

HAB/HAD Force Feedback: Module M1 extended with hip abductor force feedback
Frontal Plane Balance: Module M3 extended for lateral trunk control using θ_f
Frontal Plane Coordination: Module M4 extended for contralateral frontal plane effects
Frontal Plane Swing Control: Module M6 extended with α_f target tracking

Supraspinal Control Layer

The supraspinal layer provides:

Foot placement targets: Calculated from COM position and velocity relative to stance foot
Swing leg selection: During double support, selects leg farther from target for swing
Target angle computation: The target frontal angle is calculated by adjusting the global frontal angle based on the leg side (right or left).

Implementation Details

Control Parameters

2D control: 51 parameters (reflex control gains and targets)
3D control: 63 parameters (includes frontal plane hip abductor/adductor control)
Additional pose parameters: 26 parameters (8 joint angles + 18 initial muscle activations for starting pose)
Total without exoskeleton: 77 parameters (2D) or 97 parameters (3D)

Parameter categories:

Target angles: theta_tgt, knee_tgt, ankle_tgt, mtp_tgt
Foot placement: alpha_0, C_d, C_v (and frontal plane equivalents)
Module gains: Muscle-specific gains for each reflex module
Thresholds: Phase transition and co-activation thresholds

Sensory Processing

The system processes sensory inputs with biologically realistic delays:

Joint angles and velocities: Hip, knee, ankle, MTP
Muscle forces: All major leg muscles (GLU, HAM, VAS, GAS, SOL, etc.)
Ground contact: Binary contact sensors with medium delays
Load sensors: Continuous force sensors for stance modulation

Delay Structure

Short delays (3ms): Hip muscles, supraspinal commands
Medium delays (5ms): Knee muscles, contact sensors
Long delays (10ms): Ankle muscles, force sensors
Long loop delays (15ms): Supraspinal processing

Phase Detection

Automatic phase detection based on:

Ground contact: Binary foot-ground contact
Load thresholds: Stance initiation/termination
Swing initiation: Supraspinal swing selection during double support
Sub-phase transitions: Sw1-Sw4 trigger events

References

Song, S., & Geyer, H. (2015). A neural circuitry that emphasizes spinal feedback generates diverse behaviours of human locomotion. The Journal of physiology, 593(16), 3493-3511. DOI: 10.1113/JP270228
OpenSim-RL Implementation: osim-rl/loco_reflex_song2019.py

Usage

The reflex controller is integrated into the MyoAssist environment:

2D control: 51 parameters, sagittal plane only
3D control: 63 parameters, includes frontal plane
Delayed mode: Includes biological delays (default, requires 1ms timestep)
Non-delayed mode: Simplified for faster computation
Debug mode: Provides module-level output monitoring ```