The question of the pd-ee-pose controller.

[YZY14606]

Hello, I have a question about the PD-EE-Pose controller that I’d like to ask you. I’d really appreciate your help!

My initial end-effector pose (ee-pose) was [-2.1500e-01, 1.2178e-05, 4.9999e-01, -1.0435e-06, 3.9004e-01, 9.2080e-01, 2.0643e-07].
Expressed in Euler angles, the pose is [-2.14998186e-01, 1.21775083e-05, 4.99986112e-01, -3.14159222e+00, -2.08277065e-06, 2.34025257e+00].

I then executed the following action using base_env.step(action):
[4.17337497e-01, 1.11572049e-02, 4.94152635e-01, -3.14173972e+00, -5.37160889e-04, 2.34051285e+00, -1.00000000e+00].

After this step, the ee-pose became [-0.2048, 0.0039, 0.4987, 3.138705385175622, 0.004494347194997369, -2.9391531911425686].
As can be seen, the signs of the roll and yaw components changed, and the yaw also exhibited a large change in absolute magnitude. What could be causing this behavior?

Looking forward to your reply.

Best wishes!

[StoneT2000]

Are you using the delta version? delta target version? or absolute controller?

The PD EE pose controller in a simulator (and most others) doesn't mean some desired change in xyz or orientation will be achieved by the next step. It is an optimization problem the physics engine solves, convergence is not guaranteed after just a few physics steps (in maniskill default is 5 physics steps per control step or action sent).

[YZY14606]

I'm using absolute controller.

According to the controller design, when I provide an action, the robot’s end-effector (EE) pose is expected to move toward this action. However, in practice, the EE undergoes a large change in its Euler angle about the (z)-axis, and it ends up moving away from the commanded action instead.

This is the main issue that confuses me, and I would like to understand the underlying cause or find an effective way to resolve it.