search

silvery107 / rl-mpc-locomotion

Deep RL for MPC control of Quadruped Robot Locomotion
https://docs.google.com/presentation/d/18bznpYrkCPnhCisySPDz18hvL3Ytere7JiJEbdLvpgU/edit?usp=sharing
MIT License
417 stars 47 forks source link

Small shifting with zero command #2

Closed linchangyi1 closed 1 year ago

linchangyi1 commented 1 year ago

Hi @silvery107,

Thanks a lot for releasing the nice project. During studying the code base, I met some issues and want to discuss them with you.

  1. In the update function within /MPC_Controller/common/StateEstimator.py, the current code for updating vBody and omegaBody is as follows: self.result.vBody = self.result.rBody @ self.result.vWorld self.result.omegaBody = self.result.rBody @ self.result.omegaWorld However, I believe the correct transformations should be: self.result.vBody = self.result.rBody.T @ self.result.vWorld self.result.omegaBody = self.result.rBody.T @ self.result.omegaWorld This is based on the assumption that we want to transform the linear and angular velocities from the world frame to the body frame, where v_b = R_b_to_s @ v_s?

  2. Additionally, I noticed a minor drift in the yaw-axis and the y-axis when the command is set to zero, as shown in the following video link: https://github.com/silvery107/rl-mpc-locomotion/assets/71823391/0c03f991-0c5c-46b8-b499-c07a644657c9 Could you please give some suggestions on how to improve it?

Thanks a lot for your time and look forward to your response~~~

silvery107 commented 1 year ago

Hi Changyi,

Thanks for your interests in my project.

  1. The transpose is done within quat_to_rot() due to the rotation convention from Cheetah Software.
  2. The drift is because the MPC states I used are alined to the robot, i.e. no position constraints except height are enforced, so the drift is expected. You can use the world frame states and reference to eliminate that drift.
linchangyi1 commented 1 year ago

Got it~ Thanks a lot for your reply.