Simulation speed up

[SeaingAnt]

The simulation set with the backend np, pinocchio, etc. is quite slow when the number of cf2 grows. I was trying to solve this issue so I manipulated the attribute dt to 0.001 in order to increase the integration time and the drones became unstable when moving in xy. I tried goto, giving velocity etc., but nothing they always fall down. At first, I thought it was a problem with the integration and physical solver but when I tried to change just the dt in the step function I had no more the issue:

uav.step(actions, 0.001)#self.dt)

The time function in the backend is used by the controller to compute the tick and it is the only place where its use can explain this behaviour because, for example, the takeoff works perfectly in any case. Looking into the firmware, the tick is only used here to decide whether to run the controller or not. Considering that I tried pid and Mellinger with the same results, does anyone know why this happens?

[whoenig]

I have seen the same issues. It works with larger dt for the Lee controller (which isn't merged in the official firmware, yet), but the reason is unclear.

Some other notes on speed improvements: We are planning to add vectorization, then at least the multi-drone case should not scale purely linearly with the number of robots anymore.

[SeaingAnt]

ok, sure the vectorization should help. If it can be of any help, I also tried to remove the longest computations. In _fwsetpoint_to_sim_data_types_state I made a c code to find the quaternions, replacing the numpy one, while in setState I removed the quat to rpy conversion. In this way, I could be under 5e-4s for each iteration for 10 drones.

[SeaingAnt]

For the dt is really weird since there is no direct use of the time in the controller. Maybe this is a problem that originated from the python binding somehow.