Feasibility of Integrating a Custom MPC Controller in gym-pybullet-drones

[MohammadAlawi]

Hello, I am currently working on my thesis, which involves designing a Model Predictive Control (MPC) framework for a multi-agent drone system. The goal is to coordinate a swarm of drones to move from their respective initial positions to predefined target positions while maintaining synchronization.

I have developed and tested the MPC controller successfully using the do-mpc simulation environment. To validate the controller in a more physically realistic simulation, I integrated it into gym-pybullet-drones by modifying the pid.py controller to use my MPC logic instead of the built-in PID controller, specifically with the cf2x drone model.

However, during the simulation, the drones often fail to follow the target trajectories and frequently fall instead of hovering or moving forward. The controller appears to run and produce control inputs, but the resulting behavior is unstable or non-responsive.

My question is:

Is it feasible and supported to implement a custom MPC controller within the gym-pybullet-drones environment?

What are the common pitfalls or necessary considerations that might explain why my MPC controller fails to control the drones properly within this simulation?

Any guidance or recommendations would be greatly appreciated.

Thank you for your excellent work on this framework.

[JacopoPan]

As I wrote in #276, I would start implementing the MPC controller in the single agent environment of https://github.com/utiasDSL/safe-control-gym and only later bringing it back to gym-pybullet-drones (in principle it should work but what is the control input of your controller? CTBR? how do you handle the motor allocation?)

[wzjoriv]

Did you take into account the control frequency? And how long it takes to obtain an solution to the optimization? Because you must be able to respond quick to disturbances? MPC can be slow and by the time a control is computed the system may have changed drastically