Abstract
Tendon drives paired with soft muscle actuation enable faster and safer robots while potentially accelerating skill acquisition. Still, these systems are rarely used in practice due to inherent nonlinearities, friction, and hysteresis, which complicate modeling and control. So far, these challenges have hindered policy transfer from simulation to real systems. To bridge this gap, we propose a sim-to-real pipeline that learns a neural network model of this complex actuation and leverages established rigid body simulation for the arm dynamics and interactions with the environment. Our method, called Generalized Actuator Network (GenAN), enables actuation model identification across a wide range of robots by learning directly from joint position trajectories rather than requiring torque sensors. Using GenAN on PAMY2, a tendon-driven robot powered by pneumatic artificial muscles, we successfully deploy dynamic but precise goal-reaching, ball-in-a-cup, and table tennis policies, trained entirely in simulation. To the best of our knowledge, this result constitutes the first successful sim-to-real transfer for a four-degrees-of-freedom muscle-actuated robot arm.
How GenAN Works
Learn the hard part (the muscles and tendons) from data, and simulate the easy part (the rigid arm and its environment) with a physics engine.
Train GenAN
GenAN is a learned actuator model for robots without torque sensors. It learns the mapping from control signals to resulting joint torques with a differentiable position-matching loss.
Train the policy in simulation
The frozen GenAN converts the policy's control signals into torques for a GPU-parallelized MuJoCo XLA simulation. An ensemble of GenANs keeps the policy away from regions where the model is uncertain.
Transfer to the real robot
The trained policy is deployed on the physical robot zero-shot, without any fine-tuning on real-world data.
The Robot: PAMY2
A 4-DoF tendon-driven arm in which each joint is driven by an antagonistic pair of pneumatic artificial muscles. These muscles are fast, lightweight and inherently compliant, but their nonlinearities, friction and hysteresis make them notoriously hard to model.
Zero-Shot Transfer Results
Success rates of policies trained purely in simulation and evaluated on the real robot (100 trials each).
BibTeX
@inproceedings{schneider2026sim,
title={Sim-to-Real Transfer for Muscle-Actuated Robots via Generalized Actuator Networks},
author={Schneider, Jan and Mahajan, Mridul and Chen, Le and Guist, Simon and Sch{\"o}lkopf, Bernhard and Posner, Ingmar and B{\"u}chler, Dieter},
booktitle={Conference on Robot Learning (CoRL)},
year={2026}
}
