Is it possible to deploy this controller to a real Unitree G1 robot?

[uzakotim]

Is your feature request related to a problem? Please describe.
I have been thinking about adapting this controller to a simulated g1, by running unitree_mujoco separately, like in low-level examples from Unitree

The controller is SO good! I really appreciate your creation and thanks in advance for solving my request!

[manumerous]

Hi Timur! I am glad to hear that!

I am not sure I understand your feature request fully, could you specify in more detail what you would like? Are you planning to switch out the rigid body computations from pinocchio + CppAD (I know the fact that you need to compile the first time is a bit annoying, but it's blazingly fast) with the rigid body dynamics from Mujoco? That sounds like a major undertaking and is not something that I will find time for (I assume performance will be similar if you use CppAD for the gradients).

We already have a G1 as an example (more robotis to follow soon. Hopefully Neo once 1X shares the urdf). Do you have a link to what you have in mind?

In any case I am glad you find the work useful and happy to see more people contribute to and adapt our work!

[uzakotim]

Hello Manuel,

I would like to read and redirect joint states from a real robot through "rt/lowstate" DDS topic to the mpc and send back commands from mpc to the robot, thus making it possible to deploy the controller. The goal is to have a keyboard controller (I have done that) that publishes to "/humanoid/walking_velocity_command" and the mpc to publish to "rt/lowcmd"

It is probably difficult to do, maybe I do not understand the workflow of the controller.

[manumerous]

Ah I understand now. That should not be too difficult to do. You can just simply use the "CentroidalMpcMrtJointController" (or the Whole body version if you prefer) and integrate that into your controller pipeline. the class (and the BaseController it is derived from have a simple interface:

  // Computes [commands] given [joints] states and configs.
  virtual void computeJointControlAction(scalar_t time,
                                         const ::robot::model::RobotState& robotState,
                                         ::robot::model::RobotJointAction& robotJointAction) = 0;

The only thing you need to do is populate the RobotState object correctly from "rt/lowstate" and make sure the RobotJointAction is send correctly to the unitree API. You should be able to understand both of those classes from their implementation in the corresponding files. They are just containers that allow you to acces the right joint with the defined joint_index.

[manumerous]

You can take a look how this is handled in the mujoco sim interface and you can probably reuse a lot of the code there. Just make sure that you replace the sim interface with your ros pub/sub architecture to communicate to the robot.

humanoid_nmpc/humanoid_centroidal_mpc_ros2/src/CentroidalMpcRobotSim.cpp

Let me know if there are any further questions.