How to Solve the Issue of Continuous Control in a Robotic Hand with MuJoCo

[binbinbai]

Hi,

I'm a student and I'm trying to use MuJoCo for a robotic hand grasping task .

For this robotic hand grasping task that involves a sequence of joint angles q_tar with a length of 100, where each frame represents the desired target joint angles (including 3 translational and 3 rotational joints), I encountered an issue during the control process.When I use

for i in range(100):
    data.qpos = q_tar[i]
    mujoco.mj_forward(model, data)

the grasping sequence is accurate,the manipulator can grasp object visually.

However, this approach does not account for collisions or other physical interactions, so the object is not actually grasped. To achieve an actual grasp, I then used

for i in range(100):
    data.ctrl = q_tar[i]
    mujoco.mj_step(model, data)

but I found that this control method resulted in significant discrepancies.

Here is my actuator configuration:

   <actuator>
        <general name="A_ARTx" joint="ARTx" ctrlrange="-2 2" biastype="affine" gainprm="500 0 0" biasprm="0 -500 0" />
        <general name="A_ARTy" joint="ARTy" ctrlrange="-2 2" biastype="affine" gainprm="500 0 0" biasprm="0 -500 0" />
        <general name="A_ARTz" joint="ARTz" ctrlrange="-2 2" biastype="affine" gainprm="500 0 0" biasprm="0 -500 0" />
        <general name="A_ARRx" joint="ARRx" ctrlrange="-3.14 3.14" biastype="affine" gainprm="500 0 0" biasprm="0 -500 0" />
        <general name="A_ARRy" joint="ARRy" ctrlrange="-3.14 3.14" biastype="affine" gainprm="500 0 0" biasprm="0 -500 0" />
        <general name="A_ARRz" joint="ARRz" ctrlrange="-3.14 3.14" biastype="affine" gainprm="500 0 0" biasprm="0 -500 0" />
    </actuator>

It's important to note that the joint angle differences between consecutive frames in the q_tar sequence are not uniform, with differences ranging from 0.001 to 0.1. I think that when the difference between consecutive frames is very small (0.001), the resulting control signal is also very small, making it difficult to achieve precise control.

How can I solve this issue of continuous control?

If frame-by-frame control is difficult to achieve, is there a feasible method to smoothly move the robotic hand to the joint angles of the last frame, only ensuring that the final frame's joint angles are accurate?

[Balint-H]

Hello! Can you explain a bit better the discrepancy you observe between the two cases? The actual hand shape looks similar, it just seems to be offset.

[binbinbai]

Hello, thank you for your reply. The problem I want to resolve is the offset shown in the images.The hand shape is fine. @Balint-H