Imperial College London · 2024–2025

Markerless Pose Feedback for RL-Controlled FES

A camera-based movement feedback layer for a pre-existing reinforcement-learning-controlled functional electrical stimulation system.

Contribution: Pose estimation · system integration
Context: Master’s research project
Supervisor: Prof. Aldo Faisal
Stack: Python · BlazePose · FES

Feedback loop connecting a target, reinforcement-learning agent, stimulator, user movement and pose-estimation model — The system closes the loop by feeding observed movement back to the controller.

Pose models reviewed: 10
Camera pipeline: ≈25
Outside-lab study: 4
Prototype setup: 5–10

01 · The challenge

System objective and control constraints

The controller needs joint-angle feedback without the setup burden of a marker-based motion-capture system.

I compared ten pose-estimation approaches, selected and adapted BlazePose, then converted its landmarks into shoulder and elbow angles using vector calculus.

The camera pipeline was integrated into the existing interface so the controller could receive frames, angles and angular velocities during operation.

02 · My contribution

Pose estimation, evaluation and integration

Perception

Adapted BlazePose for markerless shoulder and elbow tracking from an iPhone camera.

Evaluation

Tested four participants outside the lab and four participants against a goniometer reference in the lab.

Integration

Connected the movement signal to the existing RL-FES interface and a portable hardware prototype.

Evaluation snapshot

The markerless pipeline held approximately 25 FPS outside the lab. Joint-angle error stayed below 3° for much of the tested range, with larger errors at biomechanical limits.

This was a small research prototype evaluated with four participants per study, not a medical device or clinical treatment system.