Arduino MyoWare EMG: Muscle Biomechanics With Arduino Tutorial

TinyML (Tiny Machine Learning) enables running optimized machine learning models on low-power microcontrollers like the Wio Terminal (ARM Cortex M4F core at 120 MHz), allowing intelligent devices to perform tasks such as gesture recognition using a single light sensor. The process involves collecting sensor data with Data Forwarder, processing it through pre-processing blocks (flatten, spectral analysis, or raw data), training a neural network classifier, and deploying the model to the device for real-time inference. The video demonstrates achieving 92.4% accuracy in rock-paper-scissors gesture classification using a 35KB neural network, illustrating how TinyML can transform simple sensors into intelligent devices.

Deploying machine learning on battery-less hardware requires addressing unique challenges. Unlike single-activation tasks, data streaming applications like gesture detection require both sufficient energy and adequate data collection before execution. The triggering mechanism becomes complex as these conditions must be met simultaneously. When energy storage saturates before sufficient data is collected, energy is wasted waiting for classification windows to complete. Training involves acquiring data from multiple participants performing labeled gestures, then dividing datasets into training, validation, and testing sets. Partnering with ML platforms enables training simple C code models using only 8 or 16-bit integers for compatibility with any embedded microcontroller. For a 2-second accelerometer window sampled at 10 Hz, the system consumes approximately 720 microjoules and takes about 30 milliseconds for processing. Sleep state with continuous data gathering consumes 65 microwatts. Cold start time from complete power-off to functional state takes approximately 10 seconds under 900 lux illumination. Practical demonstrations show gesture recognition controlling smart lights (changing color/intensity) and robotic arms through web interfaces, transmitting Bluetooth packets containing gesture classifications or inclination data.

A practical tiny ML example demonstrates gesture recognition using Arduino Nano BLE Sense and TensorFlow Lite. The process involves two phases: first, capturing accelerometer and gyroscope data from the IMU while performing gestures (punching and flexing) and saving as CSV files. Second, using Google Colab to train a TensorFlow model on this data, generating a model file compatible with TensorFlow Lite Micro. The application displays confidence percentages for detected gestures, enabling an emoji keyboard where gestures trigger corresponding emojis. This demonstrates the complete workflow from data capture to on-device inference.

Tiny ML experiments leverage web technologies through JavaScript and Web Bluetooth API, enabling anyone with a modern browser to interact with microcontrollers. Practical demonstrations include playing virtual drums based on accelerometer gestures and controlling web browsers using finger movements. The system establishes a two-way API where websites send trained models to microcontrollers and receive real-time sensor data. This approach eliminates complex driver requirements and makes ML accessible through familiar web interfaces, demonstrating how tiny ML transforms abstract concepts into tangible interactive experiences.

Gesture recognition on Arduino uses IMU (accelerometer, gyroscope, magnetometer) to detect movement patterns through data capture, visualization, model training in Google Colab with TensorFlow, and conversion to Arduino .h format. Object classification with OpenMV and Edge Impulse involves collecting images, organizing by class, uploading to Edge Impulse, configuring model parameters, generating features, and training the algorithm. The Pixel Automl platform supports Nano BLE Sense with optimized classifiers for Cortex-M0 boards. Arduino collaborates with Microsoft (Visual Studio Code extension, Azure cloud) and Google (TensorFlow Lite Micro, Coral accelerators), supporting multiple frameworks beyond TensorFlow including simplified algorithms and lightweight AI libraries designed for microcontrollers.

This comprehensive section introduces the field of biomimetic prosthetic design and traces the development of the CIS prosthetic. The human musculoskeletal system consists of bones activated by monoarticular and biarticular muscles, controlled by the central nervous system. In amputees, this system is partially lost and must be replaced by mechanical systems. Lower limb prosthetics include knee and foot components that replace bone and muscle functions, but are passive and cannot enable active joint movements. Current devices struggle with diverse locomotion conditions, leading to fall risks exceeding 50% in amputees. Since the 1970s, robotic prosthetics have been developed but face trade-offs between agility and energy efficiency. The CIS project began in 2011 with ANR funding, collecting data from 100 amputees across five daily life scenarios. In 2014, a microprocessor-controlled prototype with biarticular architecture was proposed, enabling synergistic behavior between knee and ankle joints controlled by a single hydraulic unit.

Biomedical engineers design and develop medical devices by combining engineering principles with biological systems, such as creating prosthetic sockets that integrate myoelectric sensors to enable amputees to control artificial limbs through natural muscle activity signals.

Bionic prosthetics are advanced artificial limbs that use electronic signals from the user's muscles (via electromyography sensors) to control robotic hands and arms, enabling more natural and functional movements compared to traditional biomechanical prosthetics that rely on mechanical filaments or cables; the design process involves selecting appropriate mechanisms (such as servomotors, shape memory alloys, or linkage systems), choosing suitable 3D printing materials (ABS, PETG, resin, or metal), creating custom sockets through molding or scanning, and adapting the device to each patient's specific amputation level and functional needs through a process that typically takes 1-2 years for initial development and 3-6 months for patient-specific adaptation.

This comprehensive tutorial covers the complete design and assembly of a myoelectric prosthetic circuit. The project begins with creating a symmetrical power supply using two 9V batteries connected in series, producing positive, ground, and negative terminals. The Arduino Nano microcontroller serves as the central processing unit, powered by a separate 9V battery with power distributed from both ends of the protoboard for stability. The Muscle Sensor V3 detects and processes muscle signals, filtering and amplifying electrical signals from muscle contractions. The sensor connects to the symmetrical power supply and outputs processed signals to an analog input pin on the Arduino. For actuation, the MG995 or MG996R servomotor is selected for its superior torque compared to the SG90. The servomotor requires PWM signals from Arduino pins (3, 5, 6, or 9) for precise control and operates on 6V-7.2V DC, achievable through four 1.5V batteries in series or a 5V USB charger. This complete circuit enables muscle-controlled prosthetic movement through integrated power, sensing, and actuation systems.

Biomedical engineering applies engineering principles to biological systems for medical purposes, combining engineering disciplines with biological systems to develop health technologies. Key developments include artificial organs, X-ray, electrocardiogram, cardiac pacemakers, ultrasonic imaging, MRI, and genomic analysis. Foundational disciplines include bioinformatics, biomaterials, biomechanics, medical devices, tissue engineering, rehabilitation engineering, biomedical optics, and neural engineering. Biomechatronics specifically combines mechanical, electronic, and biological systems to create health and rehabilitation products. Prosthetic development follows a systematic framework: framework development determining specifications, mechanical and biomechanical testing, conceptual design and simulation, material selection, mechanical testing following ISO 10328 standards, biomechanical testing at laboratory scale and gait analysis, and field testing with actual patients. For prosthetic foot design, criteria include lightweight construction, ability to support 100 kg, multi-axis ankle system, and normal walking capability.

Multi-channel EMG uses multiple electrode pairs placed along muscle fibers to capture spatial distribution of muscle activation. By placing electrodes on both sides of motor unit innervation zones and taking differences between channels, researchers can identify specific regions of muscle activation. This spatial mapping reveals which parts of a muscle are being recruited during different movements or conditions.

Advanced EMG systems can incorporate multiple sensors (such as 28 sensors) to provide comprehensive muscle analysis. Each sensor is strategically placed on specific muscle groups to capture detailed electrical activity data. This multi-sensor approach enables: (1) simultaneous monitoring of multiple muscle groups, (2) more accurate assessment of muscle coordination and balance, and (3) detailed mapping of muscle activation patterns during complex movements.

Two-channel animations enable simultaneous training of separate muscle groups by inhibiting one channel while contracting another, such as isolating arm elevation without shoulder involvement. Balance animations like seesaw and water glass require maintaining equilibrium between channels, demonstrating how small activation differences produce measurable effects. These exercises develop fine-grained muscle control and inter-channel coordination essential for complex motor tasks.

One approach to improving prosthetic control is using multiple EMG electrodes (4, 6, 8, up to 64 channels) to cover the entire forearm. This allows recognition of individual finger movements. However, this approach faces challenges with generalization, as different algorithms may be needed for each user.

Multi-channel EMG systems require synchronization of signals from multiple electrodes to enable accurate analysis of muscle activation patterns. Synchronization ensures that signals from different channels are temporally aligned, allowing for proper analysis of the timing and sequence of muscle activation. This is particularly important for applications involving multiple muscles or for analysis of complex movement patterns where precise timing information is essential.