Open, Accurate, and Calibration-Free Muscle-Computer Interfaces

Paper Title

Open, Accurate, and Calibration-Free Muscle-Computer Interfaces

Publication Info

• Topic area: Muscle-computer interfaces (MCIs) for gesture recognition and real-time interaction.
• Keywords: Muscle-computer interfaces, electromyography, calibration-free, gesture recognition, real-time interaction, open-source, Myo Armband, HCI, machine learning, wearable devices.

Background and Problem

• Problem / challenge: MCIs have historically required user-specific calibration data to achieve accurate gesture recognition, limiting their generalizability and usability. Recent advances have demonstrated calibration-free MCIs but rely on proprietary datasets and hardware, restricting accessibility and replication.
• Significance: Calibration-free MCIs could enable ubiquitous, hands-free, and natural input for interactive systems, but their development has been hindered by the lack of open resources and scalable solutions.
• Motivation and related work: Prior work demonstrated calibration-free MCIs using large proprietary datasets and custom hardware, achieving high accuracy in real-time tasks. However, these solutions are not replicable due to closed resources. This paper seeks to address this gap by leveraging open-source tools, public datasets, and commodity hardware.

Solution

• Proposed approach: Development of calibration-free MCIs using open-source software, publicly available datasets, and the Myo Armband, a widely adopted reference device.
• Novelty:
  1. Demonstration of calibration-free MCIs that generalize to unseen users for real-time tasks using open resources.
  2. Contribution of open-source models, code, and evaluation environments for replication and extension.
• Procedure and key techniques:
  • Utilized the EMG-EPN612 dataset (612 participants) for training foundational models.
  • Developed two real-time interaction tasks: (1) 1D cursor control and (2) discrete gesture recognition.
  • Designed task-specific machine learning models: a continuous model for cursor control and a discrete model for gesture recognition.
  • Evaluated the models with 20 participants for Task 1 and 15 participants for Task 2, without collecting any user-specific calibration data.

Results

• Concrete findings:
  • Task 1 (1D cursor control): Mean acquisition time improved from 1.63 s (Block 1) to 1.15 s (Block 5), comparable to prior closed-source benchmarks.
  • Task 2 (Discrete gesture recognition): Response time decreased from 1.53 s (Block 1) to 1.03 s (Block 4), with an error rate of 2.4% in the final block.
  • Offline classification accuracies: 93.7% for the continuous model and 94.5% for the discrete model.
• Advantage over baselines:
  • Performance comparable to proprietary systems, achieved using open-source tools and public datasets.
  • Participants reached proficiency within 25–50 trials (~2–3 minutes), demonstrating learnability.
• Experiments / evaluation:
  • Task 1: Participants controlled a cursor using wrist flexion/extension gestures to acquire targets.
  • Task 2: Participants performed discrete gestures to dismiss prompts.
  • Metrics included acquisition time, response time, and error rate.
• Limitations and future work:
  • Evaluations were conducted in controlled settings, limiting generalizability to real-world scenarios.
  • The Myo Armband, while useful as a reference device, is no longer commercially available.
  • Future work should focus on cross-device generalization, richer feedback mechanisms, and addressing false activations in uncontrolled environments.

Summary

This paper demonstrates that calibration-free MCIs can achieve real-time gesture recognition performance comparable to proprietary systems using open-source tools, public datasets, and the Myo Armband. Two interactive tasks—1D cursor control and discrete gesture recognition—were developed and evaluated, showing strong generalization to unseen users without calibration. The work contributes open models, code, and evaluation environments, lowering the barrier for replication and extension. While challenges remain for real-world deployment, this study marks a significant step toward democratizing MCI research and enabling broader adoption in interactive systems.