Title: Achieving Sim-to-Real Transfer in Muscle-Actuated Robots Through Generalized Actuator Networks

Abstract:

The combination of tendon drives and soft muscle actuation offers the promise of robots that are not only safer and faster but also capable of acquiring skills more rapidly. However, practical implementation remains scarce because inherent nonlinearities, friction, and hysteresis make modeling and control notoriously difficult. These complexities have historically prevented the successful transfer of policies from simulation to physical systems.

To address this disconnect, we introduce a sim-to-real pipeline that utilizes a neural network to model the intricate actuation dynamics, while relying on standard rigid body simulation for arm kinematics and environmental interactions. We term this approach the Generalized Actuator Network (GenAN). GenAN facilitates actuation model identification across diverse robotic platforms by learning directly from joint position trajectories, thereby eliminating the need for torque sensors.

We validated our method using PAMY2, a four-degrees-of-freedom robot driven by pneumatic artificial muscles and tendon mechanisms. By training entirely in simulation, we successfully deployed dynamic yet precise policies for goal-reaching, ball-in-a-cup, and table tennis tasks on the physical robot. To our knowledge, this marks the first successful demonstration of sim-to-real transfer for a muscle-actuated robot arm with four degrees of freedom.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC