[Question] Frequency about gym-pybullet-drones

[jacopo-panerati]

What I’d say you are doing is to make the control frequency coarser than the physics engine stepping frequency, and yes, that is done with aggregate_physics_steps

E.g. if PyBullet freq is 240Hz, with aggregate_physics_steps set to 5 means controlling the quadrotor at 48Hz

[JacopoPan]

Hi @CoolCoder54323,

argument freq sets self.SIM_FREQ which is the inverse of the timestep PyBullet advances the simulation at each call to p.stepSimulation().

aggregate_phy_steps simply determines how many p.stepSimulation() happen within each call to env.step() (it saves a number of read/writes and some wall-clock, it is also useful if you want more accuracy in the simulation, with a higher PyBullet frequency but a fixed or coarser control step frequency).

[CoolCoder54323]

So shouldn't the simulation slow down if you set freq to 1

[JacopoPan]

No, setting freq=1 simply means that every call to p.stepSimulation() (one per env.step() iff aggregate_phy_steps==1) will advance the simulation, in the simulation's clock, by 1 second. This can happen faster that 1 second in wall clock (the one of the world you live in).

The idea is to run simulations faster than real-time, to collect more data, you can check PyBullet documentation for using it as a gaming engine, with (best effort) real-time stepping, but this is not the goal in this repo.

In general, freq=1 is not a good idea because it would result in poor physics accuracy for fast dynamics/collisions.

[CoolCoder54323]

What I was trying to do is make each action of the AI last longer in the simulation. So if I'm understanding this correctly I would do that using aggregate_phy_steps so it updates p.stepSimulation() more than once using the same action.