Upper-limb motor impairments can make everyday actions, such as reaching, grasping, or manipulating objects, extremely difficult. Assistive robotic arms offer a promising way to restore part of this functional autonomy. However, for these technologies to be truly useful, they must be controlled in a way that is intuitive, reliable, and responsive. 

The challenge is not only to design a robot capable of moving, but to design an interface that people can actually use. Traditional interfaces, such as joysticks or visual control systems, can be effective in some cases, but they may not be suitable for users with severe motor limitations. Physiological signals, such as muscle activity, provide a more direct way to access the user’s intention. 

In this project we developed a complete real-time pipeline that transforms forearm muscle activity into control commands for an assistive robotic arm. The system records sEMG signals from four electrodes placed on the forearm. These signals are filtered in real time to reduce noise and preserve the frequency range associated with muscle activity. The signal is then divided into shorter windows. Each window is normalized and analyzed by a one-dimensional convolutional neural network, which classifies the user’s gesture. The final system uses a threshold-based detector to identify muscle activation, followed by a five-gesture classifier. Each recognized gesture is mapped to a robotic command, allowing the arm to move upward, downward, laterally, or activate the gripper.