Model output to Crazyflie Control

[zcase]

Hi great repo! Do you have an example using the stablebaselines3 algorithms such as PPO that controls an actual crazyflie drone? I just trained a model using the (vp, vq, vr, T) flight mode and would like to integrate into a crazyflie. I was just curious if you already had a working example.

[jjshoots]

Hi! Thanks for dropping by.

I don't have an end-to-end sim-train-to-real example, but I have code that shows how to interact with the crazyflie using setpoints here.

The code is fairly undocumented as of now, so feel free to leave this issue open and leave questions if you want.

[zcase]

ok thank you. Quick question about you the values used to train models. are all those in the global frame or body frame?

Digging into the code found here:

    def update_state(self) -> None:
        """Updates the current state of the UAV.

        This includes: ang_vel, ang_pos, lin_vel, lin_pos.
        """
        lin_pos, ang_pos = self.p.getBasePositionAndOrientation(self.Id)
        lin_vel, ang_vel = self.p.getBaseVelocity(self.Id)

        # express vels in local frame
        rotation = np.array(self.p.getMatrixFromQuaternion(ang_pos)).reshape(3, 3).T
        lin_vel = np.matmul(rotation, lin_vel)
        ang_vel = np.matmul(rotation, ang_vel)

        # ang_pos in euler form
        ang_pos = self.p.getEulerFromQuaternion(ang_pos)

        # update the main body
        self.body.state_update(rotation)

        # create the state
        self.state = np.stack([ang_vel, ang_pos, lin_vel, lin_pos], axis=0)

        # update auxiliary information
        self.aux_state = self.motors.get_states()

it looks like the following are expressed in world/global coordinates:

• lin_pos
• ang_pos

And that the following are expressed in the body/local coordinates:

• lin_vel
• ang_vel

Is this correct?

Also all of these are in radians correct?

[jet-sony]

Yep, both of those assumptions are correct. Positions are always in global while velocities are local.