Implementation of a Surface Electromyography-Based Upper Extremity Exoskeleton Controller Using Learning from Demonstration

Ho Chit Siu, Ana M. Arenas, Tingxiao Sun, Leia A. StirlingSensors 2018 · DOI

TL;DR

This paper introduces a novel approach to controlling upper-extremity exoskeletons using surface electromyography (sEMG) and "learning from demonstration." By training the system on user interactions, it enables anticipatory control, allowing the exoskeleton to act before the user's physical input is detected, leading to more fluid and intuitive human-robot collaboration.

What Problem Does This Paper Address?

Upper-extremity exoskeletons hold immense promise for augmenting human capabilities, assisting individuals with disabilities, and aiding in rehabilitation. However, controlling these devices effectively, especially for the complex movements of the upper limbs, has been a significant challenge. Traditional control methods often rely on reactive inputs like switches or force sensors, leading to noticeable delays and a less natural user experience. While sEMG offers a more direct interface with muscle signals, enabling potentially anticipatory control, existing methods typically require precise sensor placement and can be sensitive to muscle shifts. A key paradox is that a constrained limb might not produce the expected sEMG signals, yet the exoskeleton needs these signals to learn how to move appropriately. This paper tackles the problem of creating a personalized, robust, and anticipatory sEMG-based control system for upper-extremity exoskeletons that overcomes these limitations.

Methods

The core of this work is an adaptive sEMG-based exoskeleton controller that leverages "learning from demonstration" (LfD). Instead of relying on pre-defined muscle mappings or complex system identification, the system learns a personalized mapping between muscle activation (sEMG) and desired system states through direct user interaction.

Here's a breakdown of the key methods:

  • Learning from Demonstration (LfD): The system learns by observing a human user performing a task. In this case, the user interacts with a thumb exoskeleton, and their sEMG signals are recorded alongside their physical inputs (pressure).
  • Initial Limit Switch Control: To facilitate the learning process and avoid the control paradox, the exoskeleton is initially controlled by a simple pressure-based limit switch. This allows the user to perform the task and generate labeled sEMG data without the system needing to understand their intent beforehand.
  • sEMG Feature Extraction: A rich set of 14 sEMG features are extracted from 125 ms epochs of data, providing 84 real-valued control inputs per epoch. These features capture various aspects of muscle activity, including frequency and amplitude characteristics.
  • Supervised K-Nearest Neighbors (KNN) Classifier: A supervised KNN classifier is used to learn the mapping between the extracted sEMG features and the user's intent (grasp or release). This classifier is chosen for its ability to work with small datasets and its lack of model training time after data collection, which is crucial for human-in-the-loop systems.
  • Time-Shifting Labels for Anticipation: To enable anticipatory control, the sEMG labels are shifted backward in time by one epoch. This means the system learns to associate sEMG signals preceding a physical action with that action, allowing it to predict and initiate movement before the user's physical input crosses a threshold.
  • Adaptive Controller (AC): The adaptive controller updates its sEMG mapping between trials using all collected data. It constructs a reduced model that is close to the full model, allowing for real-time adaptation and continuous improvement over the experiment. This addresses potential changes in user intent or physiological state (like fatigue).
  • Novice Sensor Placement Robustness: A key innovation is the system's robustness to non-specific sEMG sensor placement. Sensors were placed on the forearm without requiring expert knowledge of specific muscle locations, making the system more practical for real-world use.

Datasets

The study utilized data collected from 18 human participants performing a tabletop book-shelving task. This involved moving books between a tabletop and six shelves using a custom-built thumb exoskeleton. The data collected included:

  • sEMG signals: From six Delsys Bagnoli double-differential sEMG sensors placed on the forearm.
  • Force-sensitive resistor (FSR) pressure data: From sensors on the thumb to detect grasp and release intent.
  • Exoskeleton actuation data: Recording the movement of the thumb exoskeleton.

This study did not use pre-existing datasets but rather generated its own benchmark through controlled human experiments.

Key Insights and Ideas

  • Anticipatory Control is Achievable with LfD: The most significant insight is that learning from demonstration, combined with sEMG, can effectively enable anticipatory control in exoskeletons, moving beyond reactive systems.
  • Robustness to Novice Sensor Placement: The system's ability to perform well even with non-expert sensor placement is a major step towards practical, user-friendly exoskeletons. This removes a significant barrier to adoption.
  • Adaptation to User Changes: The adaptive controller's ability to update its mapping over time suggests potential for long-term usability, accommodating user fatigue or subtle changes in movement patterns.
  • Bridging the Control Paradox: The LfD approach, starting with a limit switch, elegantly bypasses the challenge of training an exoskeleton when the user's limb is constrained.
  • Short Training Times are Feasible: The system achieved effective control with relatively short training periods (around one minute), which is crucial for user acceptance and practical deployment.

Key Results

  • Anticipation Time: The NA and AC controllers demonstrated anticipatory behavior, triggering grasp actions approximately 0.19-0.29 seconds before the pressure threshold was met. This is a significant improvement over the limit switch controller, which is purely reactive.
  • Training Time: The training time for all controllers was comparable, averaging around 51 ± 14 seconds, indicating that the LfD approach does not impose an excessive training burden.
  • Precision: While the limit switch controller had no false positives (by definition), the NA and AC controllers showed comparable numbers of true and false positives. However, the duration of true positives was significantly longer than false positives, suggesting that while false positives occurred, they were often brief and less impactful on the task.
  • Multi-Grasps: The study identified and analyzed "multi-grasps," which were instances of multiple grasps within a single close-open cycle. While some were outliers, many were attributed to quick re-grasping actions during the task, suggesting potential for further refinement in instruction or controller adaptation.

Limitations and Future Work

  • Discrete Control States: The current system operates with discrete "open" and "close" states. Future work could extend this to a continuous control system for more nuanced movements.
  • False Positives: The NA and AC controllers exhibited a notable number of false positives, although they were generally short-lived. The authors suggest that these might be mitigated by refining the training procedure (e.g., by better handling of tapping motions) or by adjusting the AC update assumptions.
  • Adaptive Controller Update Rate: The adaptive controller's update rate (every few minutes) might have made it harder for users to learn compared to a static model. Further research is needed to find an optimal update rate.
  • Hybrid Control: The authors suggest that a hybrid approach combining sEMG and pressure sensing could further improve prediction performance and robustness.
  • Generalizability to Complex Tasks/Users: While the framework is generalizable, the current study focused on a specific book-placing task. Further validation is needed for more complex tasks and for users with different physiological conditions (e.g., rehabilitation scenarios).
  • Machine Learning Algorithm: The study used KNN, but exploring other algorithms like Support Vector Machines (SVMs) could be beneficial, especially for continuous updating scenarios.

Summary updated

BibTeX
@article{siu2018implementation,
  author = {Ho Chit Siu and Ana M. Arenas and Tingxiao Sun and Leia A. Stirling},
  title = {{Implementation of a Surface Electromyography-Based Upper Extremity Exoskeleton Controller Using Learning from Demonstration}},
  journal = {Sensors},
  year = {2018},
  doi = {10.3390/s18020467},
}
Figure