Performance

With modern hardware (2.2 GHz AMD Opteron) the direct control typically worked in real-time with the cart-pole simulation. On average, the traditional NMPC method was about 20 times slower than real-time and OIC more than 100 times slower. The bad performance of OIC resulted from the Kalman smoothing (see Section II-A) not converging and having to switch to the slow update mode. Further optimisations to the algorithms or improvements in hardware are clearly required, before systems with fast dynamics can be controlled.

Figure 5: Example of a successful swing-up with NMPC and high noise. The system is plotted with the observation noise included.