Sim-to-Real Transfer for Muscle-Actuated Robots via Generalized Actuator Networks
Researchers have developed a new pipeline for transferring robot control policies from simulation to real-world muscle-actuated robots. This method, called Generalized Actuator Network (GenAN), uses neural networks to model the complex, non-linear dynamics of pneumatic artificial muscles, overcoming challenges like friction and hysteresis that previously hindered sim-to-real transfer. The system was successfully demonstrated on a four-degree-of-freedom robot arm, enabling precise execution of tasks such as goal-reaching and table tennis, trained entirely in simulation. AI
IMPACT Enables more complex robotic tasks by improving sim-to-real transfer for muscle-actuated systems.