quadcopter sim2real discuss

[shaoxiang]

Is there a sim2real possibility for the quadcopter example shown in the Isaac Lab?
I noticed that the total force and torque are used to control the drone in Isaac Lab.

def _pre_physics_step(self, actions: torch.Tensor):
    self._actions = actions.clone().clamp(-1.0, 1.0)
    self._thrust[:, 0, 1] = self.cfg.thrust_to_weight * self._robot_weight * (self._actions[:, 0] + 1.0) / 2.0
    self._moment[:, 0, :] = self.cfg.moment_scale * self._actions[:, 1:]

def _apply_action(self):
    self._robot.set_external_force_and_torque(self._thrust, self._moment, body_ids=self._body_id)

However, when we use traditional control methods, we should control the motor speed, which corresponds to the force and torque of the corresponding motor. So, for example, in the set_external_force_and_torque method in Isaac Lab, how can the total force and torque correspond to the real speed of the motors?


So is it necessary to imitate Rotors—a modular
gazebo mav simulator framework to implement Isaac Lab's quadcopter simulation?
Welcome to discuss.

[StrainFlow]

I think you could take a few approaches to adapt this for sim to real. Please keep in mind that these examples are meant to be easy to follow rather than fully prepared for enterprise application like sim-to-real. That said, they should all be adaptable to higher fidelity scenarios.

You have a few options here. You could:

1. Have a 4-dimensional action space and incorporate your own thrust physics into the isaac sim
2. Have a 4-dimensional action space that you convert into a torque and moment in the pre-physics step which are then applied in the apply-actions step.
3. Use the same actions space as the sample and then incorporate some sort of conversion in the real robot.

I would take approach 2, in particular because you could include some domain randomization on the k_f parameter to make your model robust.