This work introduces a neuromorphic joint-space controller for a four-degree-of-freedom (DoF) servo-arm robotic manipulator, implemented via a spiking neural network (SNN) on low-cost embedded hardware. Unlike traditional methods based on inverse kinematics or Cartesian-space control, the proposed architecture operates directly on servo joint angles, receiving current and target configurations to predict the next joint-space action. A hybrid dataset with over 500,000 samples was generated by expanding joint-angle logs into a larger synthetic set to comprehensively cover the robot’s workspace, including oversampling of mechanically challenging configurations. The SNN was trained via surrogate gradients (snnTorch), and a systematic grid search was used to optimize the hidden-layer size, learning rate, and membrane decay factor. The final model, featuring a 128-neuron leaky integrate-and-fire (LIF) hidden layer, achieved a mean absolute error (MAE) of 11.06° on the validation set and generated smooth, reproducible trajectories during real-robot execution. Additionally, this study includes an empirical energy comparison between the SNN controller and an artificial neural network (ANN) baseline on a Raspberry Pi under identical motion routines. The results indicate that both models consume similar average power, with the SNN exhibiting a modest increase due to temporal simulation overhead in software-based execution on standard central processing units (CPUs).

This work introduces a neuromorphic joint-space controller for the 4-DoF Serv-Arm robotic manipulator, implemented via a Spiking Neural Network (SNN) on low-cost embedded hardware. Unlike traditional methods based on inverse kinematics or Cartesian-space control, the proposed architecture operates directly on servo joint angles, receiving current and target configurations to predict the following joint-space action. A synthetic dataset with over 500,000 samples was generated to comprehensively cover the robot’s workspace, including oversampling of mechanically challenging configurations. The SNN was trained with surrogate gradients (snnTorch), and a systematic grid search optimized the hidden layer size, learning rate, and membrane decay factor. The final model, featuring a 128-neuron LIF hidden layer, achieved a Mean Absolute Error (MAE) of 11.06° on the validation set and generated smooth, reproducible trajectories during real-robot execution. Additionally, this study includes an empirical energy comparison between the SNN controller and an ANN baseline on a Raspberry Pi under identical motion routines. The results indicate that both models consume similar average powers, with the SNN showing a modest increase due to the temporal simulation overhead typical of spiking models on standard CPUs.