Help with understanding uncertified input plot for two very similar experiments

[itsLydt]

I am running some experiments based on the one in safe-control-gym/examples/cbf/cbf_experiment.sh. Here is a diff view of my two experiments;

as you can see, all I changed between the two is the safety filter setting. And here is my result graph:
cbf_nn:

cbf:

Since the controller is the same, I expected the uncertified input to be the same for both experiments, but they are quite different. And indeed, the uncertified graphs of theta vs theta_dot do look identical, just not the inputs.

Could you correct my understand on this?

[Federico-PizarroBejarano]

Hi @itsLydt, the experiment will result in different trajectories because the safety filters are different. The first command from the controllers should be the same, but since the safety filters are different they will modify the controller's actions differently, resulting in a different state trajectory and thus a different action trajectory. In the plots you shared, it is plotting the uncertified inputs and the certified inputs for the certified experiment, i.e., the uncertified input plotted there is not from the purely uncertified experiment but the attempted input of the controller during the certified experiment. Does that make sense?

[itsLydt]

Ah yes got it. Is that true for the plots of theta vs theta_dot as well?
Is there a way to somewhat easily obtain a plot of what the controller would have done had the filter not been there at all?

[Federico-PizarroBejarano]

Yes, all the current plots are from the certified results, as seen here:

safe-control-gym/examples/cbf/cbf_experiment.py

Line 83 in 0e5f4f1

ax_obs.plot(certified_results['obs'][0][:, 0], certified_results['obs'][0][:, 2], '.-', label='Certified')

. You can instead plot the uncertified results using the data from here:

safe-control-gym/examples/cbf/cbf_experiment.py

Line 49 in 0e5f4f1

results, _ = experiment.run_evaluation(n_episodes=n_episodes, n_steps=n_steps)

.