We test if muscle actuator morphology facilitates learning by applying state-of-the-art learning algorithms covering an extensive range of approaches currently used in robotics. The common thread of the selected algorithms lies in their independence of the actuator morphology: this allows us to easily exchange idealized torque actuator morphology with muscle actuator morphology. We choose optimal control, model-predictive control and reinforcement learning as learning approaches.
The implemented code using these learning approaches can be found here:
https://doi.org/10.18419/darus-3268 (for MPC and OC code)
https://github.com/martius-lab/learningwithmuscles (for RL code)