All realistic controllers are unstable

[Astik-2002]

Greetings. I'm trying the controller examples given in realistic folder, and all controllers that are working properly in ideal section are unstable in realistic setting. Is this by design, where we have to make our own controllers or modify existing one's for better results? Also, can you please share how ideal and realistic environments are different from each other?

[fwiebe]

Hi @Astik-2002 ,
thanks for our question. The ideal examples only use the pure dynamical model of the double pendulum as described here. In the realistic examples additional effects such as process and measurement noise, delay, torque noise, torque response and perturbations are also simulated (not all effects are turned on in all scripts). The effects are the effects that are tested in the robustness score (see robustness leaderboards for acrobot and pendubot ).
The controller parameters in the scripts in the realistic folder are not tuned. So there may be parameters with which the controllers are more stable. But the scripts are just supposed to be examples for testing controllers with the mentioned effects without running the whole benchmark pipeline.