Manipulator Agnostic Gripper Control for Reinforcement Learning

[vmstavens]

Intro

Hi!

I am a Ph.d student and am using MuJoCo (MJX) for research in Deep Reinforcement Learning. I am unsure rather to post this question in the Brax repo or here, so sorry if this is misplaced :)

My setup

My system setup is

• OS: Ubunt 24.04
• Python3.12
• MuJoCo: 3.3.3
• Brax: 0.12.4
• GPU: NVIDIA 5080

My question

I am looking to train an RL algorithm to move and actuate a gripper for manipulation tasks. While i could attach the gripper to a manipulator like the Franka Panda or the UR, I ideally want to train a policy which produces changes in end-effector poses such that I can handle the manipulator control externally and have the policy be manipulator agnostic.

I know that in MuJoCo there exists mocap objects which i can weld to the gripper with a free joint and then control the grippers pose using the mocap object.

Using sliding joint actuators I also have to deal with actuator dynamics and their gains which is a non issue with the mocap object, and therefore at the moment i would prefer the mocap approach. Is it possible to use the MLP outputs to control the mocap pose without actuators?

A minimal example, pseudo code or just some clarification would me immensely appreciated! :)

Thanks in advance!

Minimal model and/or code that explain my question

No response

Confirmations

• I searched the latest documentation thoroughly before posting.
• I searched previous Issues and Discussions, I am certain this has not been raised before.

[yuvaltassa]

Why aren't you using a Cartesian controller?

https://youtu.be/s-0JHanqV1A (link to model therein)

[vmstavens]

I've been exploring this approach for some time now, and I've managed to achieve a similar effect using the Cartesian controller.

Thanks you for the reference! :)