Andrew Thompson

Bias-Aware Interface Remapping for Assistive Teleoperation

Expanding access to multi-DOF robotic control by optimizing for individual motor constraints
Andrew Thompson, Larisa Loke, Brenna Argall | Northwestern University + Shirley Ryan AbilityLab

Overview

This project developed and deployed a bias-aware interface remapping algorithm for individuals with motor impairments using low-DOF joysticks. The system reinterprets sparse input patterns to expand access to the full control space of assistive robots—without relying on explicit mode switching.

Unlike fixed axis-aligned control mappings, our method statistically analyzes user-specific input behavior and generates full-rank mappings using convex hull expansion and sparse vector decoding. The system was tested in an IRB-approved study involving participants with SCI and stroke.

Motivation

  • Many users operate assistive joysticks in strongly biased patterns (e.g., circular motion or single-axis preference).
  • These constraints reduce access to the full range of robotic functionality (e.g., 6-DOF control).
  • Existing solutions rely on mode switching, which adds cognitive burden and impedes fluidity.
  • Our system remaps these biased inputs into full-rank command spaces by learning from natural user behavior.

System Design

Input Space Modeling

  • Captured joystick trajectories across calibration tasks.
  • Computed empirical convex hull of user input space.
  • Identified input directions with low variance or dropout.
  • Constructed a mapping matrix to distribute joystick movement over 6-DOF robot control space.

Mapping Strategy

  • Used sparse basis expansion to project 2–3 axis joystick data to a higher-rank output vector.
  • Added gain compensation and thresholding logic to stabilize near-degenerate directions.
  • Embedded mapping in real-time teleoperation loop using ROS2.

Deployment & Evaluation

  • Study Participants: N = 8
    (5 SCI, 3 Stroke)
  • Hardware: 3-axis assistive joystick (mounted), Kinova Gen2 robotic arm
  • Tasks:
    • 2D planar reaches
    • 3D endpoint positioning
    • Orientation alignment tasks

Metrics Collected

  • Task completion time
  • Path smoothness
  • Input command entropy
  • Control subspace utilization
  • GUI logs and qualitative feedback

Key Contributions

  • Developed a real-time remapping algorithm customized to user-specific bias patterns.
  • Enabled full 6-DOF control from a low-DOF joystick with no mode switching.
  • Designed a custom OpenGL-based GUI to collect detailed usage metrics and visualize interface behavior.
  • Demonstrated measurable performance improvement and increased expressiveness over baseline mappings.

Design Tradeoffs

Challenge Approach
Near-singular input patterns Used convex hull regularization + sparse decoding
Real-time stability Applied smoothing and bounded velocity control
Adaptability vs. user learning Fixed mapping per session to allow learning, but tunable between sessions
Interface generality Mapping tailored per user but used shared core logic

Citation

ICORR 2022:
Control Interface Remapping for Bias-Aware Assistive Teleoperation
Andrew Thompson, Larisa Loke, Brenna Argall Read the full paper (PDF)

Access

The interface mapping logic and experimental task GUI are implemented in a private ROS + OpenGL research repository.

  • Summary PDF, logs, and sanitized code excerpts available upon request.
  • Core algorithm written in Python; live integration with Kinova control via ROS2 services.