Sensorless variable impedance control of robot joints with 3K gear reducers based on hybrid dynamics model and meta-policy optimization

Abstract

Achieving both high-precision trajectory tracking and compliant physical interaction in collaborative robot joints equipped with 3K planetary gear reducers remains challenging under torque-sensorless conditions. This paper proposes a physics-informed data-driven joint dynamics modeling method and an associated meta-policy-optimized adaptive variable impedance control scheme to address this limitation. First, a hybrid dynamics model combining lumped rigid-body dynamics with a Multilayer Perceptron (MLP) for residual loss estimation is established. By incorporating power-flow direction features, the network identifies and predicts the time-varying transmission efficiency and nonlinear friction losses of the 3K reducer across motoring and generating quadrants. Second, a model-based feedforward torque compensation scheme is designed, and a variable-gain generalized momentum observer is developed to achieve stable estimation of external interaction torques. Building upon this, a force-responsive variable-stiffness impedance control law is formulated, and a proximal policy optimization (PPO) algorithm is employed within a contextual reinforcement learning architecture to search for optimal high-level meta-parameters, enabling adaptive adjustment of impedance parameters. Multi-condition dynamic interaction simulations demonstrate that the proposed hybrid modeling approach achieves a 40% improvement in joint torque prediction accuracy compared to the conventional momentum disturbance observer (MDOB) baseline. Furthermore, the adaptive variable stiffness control law actively adjusts the joint stiffness in response to estimated external forces, satisfying the dual requirements of high-precision tracking in free space and high compliance in contact space.