We present a system for automatically creating control policies for a simulated biologically based arm performing basic tasks such as reaching for an object. The arm is driven by forces from a set of Hill-type musculotendon units (MTU). Each MTU is controlled by an excitation parameter that represents the overall activation of the muscle by the motor neurons. The control policies use splines to drive simple PD-controllers which in turn generate the muscle excitation signals which drive the MTUs. Spline control points and PD-controller parameters are learned for a specific task using a reenforcement learning based optimization which rewards how well the arm completes the task and penalizes energy usage. We use this general system to generate controllers for specific tasks.
Optimized rest arm rest state
Arm reaching for several targets