Control Airplane example

[lcaquot94]

Hello there ! I would like to work with genesis/assets/meshes/Airplane/airplane.obj for a toy scenario. The goal would be to control it with roll, pitch, and yaw at constant speed, and train it to reach target points.
What would be the good may to make the plane move in the scene ? I have tried "set_pos()" method but I can not handle angles since there is no set_angles() method. This plane differs from Control your Robot example since it has no degree of freedom, so I am not sure this tutorial is valid for my use case...

[Tadinu]

Regarding qpos for a free-floating object, It actually has 6dofs assumed by a free joint.
You may wanna ref

• https://github.com/Genesis-Embodied-AI/Genesis/blob/main/examples/rigid/ik_duck.py
• euler_to_quat for rpy to quat as for the inputs

For example, as modification to ik_duck:

kinematics_only = False
if kinematics_only:
   robot.set_qpos(q)
else:
   assert robot.n_dofs == 6
   dof_idx = np.arange(6)
   ctrl = np.empty(6)
   ctrl[:3] = q[:3].cpu().numpy()
   ctrl[3:] = quat_to_xyz(q[3:].cpu().numpy())
   robot.control_dofs_position(ctrl, dofs_idx)
   scene.step()

To make it stable, you may need to fine-tune the gains like in control_your_robot.py

[duburcqa]

Yes. I will just change the default soon.

[lcaquot94]

Ok thank you for answering my question. I also wonder if genesis is designed for position / velocity control (assuming we already have a way to reach set positions/velocities) or if it is better to always refer toi forces and torques, and let the Genesis engine do the magic and calculate new positions ? Are both approaches compatible with Genesis ?

[duburcqa]

Can you be more specific? Although there is no proper actuator abstraction for now, Genesis proposes both position/velocity control or torque control at joint-level, via control_dofs_position, control_dofs_velocity and control_dofs_force. One other option is to rely on Inverse Kinematics (IK) to specify the desired pose (position + orientation) of a link at a given point in time and let the optimisation compute the path for you.

[lcaquot94]

To be more specific, I'm interested in using a function that takes as input B 3D vectors (either velocities or accelerations, where B is the batch size). These B vectors would be expressed in the object's local frame. The function would then handle updating the object's velocity, position, qpos vector, etc.

The goal is to use this for Reinforcement Learning to control the movement of an aircraft by controlling its acceleration/velocity in its own reference frame.

[duburcqa]

If you control the acceleration directly, then you don't need a simulator basically x) I would say that enforcing the position / velocity directly is even worst, because in this case there is no warranty that the position/velocity/acceleration are consistent with each over (ie they corresponds to the same underlying trajectory being integrated over time).

The only physically realistic way to interact with a physics simulator is the same as in realistic: through external efforts (linear forces or torques). The methods control_* does exactly, either by controlling external forces directly, or by using PD control.