Possibility of software-in-the-loop (SIL/SITL) testing about gym-pybullet-drones

[dominiknatter]

Thanks for the clarifications @JacopoPan!

[JacopoPan]

Yes @dominiknatter, you are correct.

The point here is the simulation, reinforcement learning, and control of dynamical systems. I chose the CF because it's a platform I like, use, and with a lot of system identification work available.

There are similarities between the handwritten PID controllers and the CF's firmware control modules but, ultimately, one should re-define action and observations spaces depending on the type of controller they want to implement and port the trained agent to the C firmware. (Of course it's something that has been at the back of my mind from the incipit but this public repo does not include any work in that direction.)

[soldierofhell]

ultimately, one should re-define action and observations spaces depending on the type of controller they want to implement and port the trained agent to the C firmware. (Of course it's something that has been at the back of my mind from the incipit but this public repo does not include any work in that direction.)

Hi @JacopoPan ,
I've got few questions regading "deploying" the learned policies into flight controller if you mind:

• This Crazyflie MCU (Cortex-M4) is not the most performant one on the market, don't you think?
• AFAIK Crazyflie firmware is in C, this could be a problem if the code is in C++?
• Are currently used RL frameworks (stable_baselines3, rllib) capable of deploying stripped down code suitable for MCU deployment? I thought the smoothed path is through TFLite and through tf_agents, although there're still some unsupported TFLite ops right now. I don't know about PyTorch, maybe through onnx and TVM?

Regards,