Neural-Augmented Modeling and Control of Hybrid Series-Parallel Elastic Actuators for Humanoid Robotics

Abstract

Humanoid robots require actuators that are adaptive, efficient, and cost-effective, yet conventional solutions such as harmonic drives often struggle to balance compliance, precision, and affordability. This thesis introduces a neural-augmented modeling and control framework built around a novel hybrid Series-Parallel Elastic Actuator (SEA-PEA). The actuator integrates a BLDC motor, a planetary gearbox, and a rotary pneumatic element, leveraging controlled gearbox backlash to switch between series and parallel elastic behaviors. Within the backlash region, it functions as a Series Elastic Actuator (SEA), providing tunable stiffness through pneumatic pressure, enhancing compliance and safety, and enabling the storage and release of mechanical energy. Outside the region, it transitions to a Parallel Elastic Actuator (PEA), delivering stiff, precise, and energy-efficient torque transmission. To capture these nonlinear dynamics, a hybrid modeling strategy is proposed: an Echo State Network (ESN) for the SEA regime and a Wiener model for the PEA regime. Experimental validation shows that this architecture reduces modeling error compared to conventional ANN approaches. Beyond modeling, the artificial neural network framework was applied to enhance control of the hybrid SEA–PEA actuator itself, showing improved tracking performance and robustness compared to traditional control approaches. Building on this foundation, the same neural modeling and control principles were extended to a 12-DOF humanoid robot, where ESN-based Computed Torque Control (CTC) enabled reliable sim-to-real transfer validated both in simulation (PyBullet) and on a physical robot prototype. Compared to PID and torque control, neural-augmented CTC achieved lower mean square error (MSE) and reduced energy consumption, with further improvements under external load when combined with online Recursive Least Squares (RLS) adaptation.