Student-Demonstration-Driven Embedded Deep Learning Framework for Real-Time Robotic Manipulator Control on Arduino-Class Platforms

Abstract

Low-cost robotic manipulators are typically based on the methods of analytical inverse kinematics and classical control techniques that demand precise kinematic modeling and have limited robustness against nonlinearities, parameter uncertainty, and underlying hardware limitations. Moreover, learning-based robotic control methods often require high-performance processors or external calculation devices and thus cannot be applied in educational and low-power settings. To overcome such shortcomings, a deep learning control framework driven by student demonstrations is introduced to predict inverse kinematics in real time on Arduino-class microcontrollers. The nonlinear map relating Cartesian end-effector pose to joint angles is trained on the publicly available Robotic Arm Dataset on Kaggle, which contains labeled kinematic trajectories generated by simulated manipulator motion. Using supervised learning, post-training INT8 quantization, and memory-aware deployment, a compact fully connected deep neural network (3-64-64-3 topology, 4,480 parameters) is trained offline. Experimental assessment via a software-based embedded emulation pipeline is characterized by high prediction accuracy, joint-wise RMSE of 0.003-0.006 rad, and mean absolute joint error of less than 0.004 rad. The quantized INT8 model achieves a memory footprint of approximately 5.2 KB with a mean inference latency of approximately 4.2 ms. The average end-effector positioning error is less than 2.5 cm compared to analytical inverse kinematics baselines. These findings demonstrate the feasibility of deep learning-based inverse kinematics on resource-constrained Arduino-class platforms, providing a scalable and cost-efficient solution for embedded automation, practical robotics education, and edge AI applications.