Learning the Unobservable: Generalized Actuator Networks and Sim-to-Real Transfer for Muscle-Actuated Robots

Introduction

In pharmacokinetic modeling, researchers face a fundamental constraint. They cannot measure drug concentration directly in every tissue, yet they must infer it from observable outputs such as blood plasma levels or urinary excretion rates. This process of inferring unobservable internal states from external measurements represents one of the core challenges in system identification. The recent work by Schneider et al. on Sim-to-Real Transfer for Muscle-Actuated Robots via Generalized Actuator Networks presents a striking parallel. Their method learns the complex internal dynamics of pneumatic artificial muscles using only joint position trajectories, bypassing the need for direct torque measurements that have historically limited similar approaches.

Muscle-actuated robotics, particularly systems employing pneumatic artificial muscles (PAMs) paired with tendon drives, offers compelling advantages over traditional motor-driven designs. These systems exhibit inherent compliance, reduced moving mass, and enhanced safety for human interaction. However, they have remained largely confined to research curiosities rather than practical deployment. The obstacle is not merely mechanical complexity but epistemological; these systems exhibit severe nonlinearities, hysteresis, and time-varying properties due to temperature fluctuations and material fatigue. Traditional physics-based models fail to capture these dynamics with sufficient fidelity for sim-to-real transfer, while domain randomization techniques require approximately accurate base models that have proven elusive for such hardware.

Decoupling Dynamics: The GeAN Architecture

The core innovation of the Generalized Actuator Network (GeAN) lies in its hybrid modeling approach. Rather than attempting to simulate the entire system from first principles or learning an entirely black-box dynamics model, Schneider et al. separate the analytically tractable components from the intractable ones. They employ established rigid body physics for the arm dynamics and environmental interactions while learning a neural network model exclusively for the actuation subsystem.

This architectural decision carries significant practical implications. Previous actuator network approaches, notably the framework introduced by Hwangbo et al. for quadruped locomotion, required torque sensors to provide training labels. These sensors, while valuable, add considerable cost, complexity, and fragility to robotic platforms. Many existing muscle-actuated robots, including the PAMY2 platform used in this study, lack such instrumentation. GeAN eliminates this dependency by formulating the learning problem as trajectory-based system identification. The network learns to predict the mapping between commanded inputs and resulting joint torques