Unusually large contact forces during trajectory replay

[chaitanya1chawla]

Intro

Hi!

I am a graduate student researcher at Carnegie Mellon University, I use MuJoCo for my research on Humanoid Manipulation.

My setup

Mujoco: 3.2.3
Python 3.8
OS: Linux Mint 22

My question

Screencast from 02-08-2025 04:39:43 PM.webm
I have a setup for teleop control of humanoid fingers and arms. I collected some pick-place demonstrations and am trying to replay them in Mujoco.
However, while replaying (by giving the commands from the recorded data), the contact forces tend to be very high, leading to the robot not opening its fingers properly (possibly due to penetration with object).

I am using a position-based controller, with -- finger kp=10, kd = 0.25

The attached video shows a recorded demonstration on the top left frames and the current scene on the bottom left frames.

I have tried adjusting kp, solimp, solref values.

Minimal model and/or code that explain my question

reproducible_code.zip

[yuvaltassa]

Are you controlling the length of the visualized forces? Otherwise it's hard to diagnose if the forces are actually "very high". If you set <statistic meanmass="1"/> then visual/map/force will be in units of Newton/Meter, which will help you understand what is happening.

[chaitanya1chawla]

Hi Yuval,
Thanks a lot for the quick response!
I set the <statistic meanmass="1"> in my XML, however, the forces still seem similar to earlier.

Here are the quantitative values:

contact 0, body1: right_wrist_yaw_link, body2: pepsi_bottle, force: [269.39612867   8.42473816  17.57145022], torque: [ -1.3602681   19.0797237  -17.41729967]
contact 1, body1: R_index_intermediate, body2: pepsi_bottle, force: [2109.47496935    3.2237303    -4.60245705], torque: [  7.50574668  31.23817751 100.56420644]
contact 2, body1: R_pinky_intermediate, body2: pepsi_bottle, force: [18.98488066  0.          0.        ], torque: [ 0.        -5.6954642  0.       ]
contact 3, body1: R_ring_intermediate, body2: pepsi_bottle, force: [19.38153308  0.          0.        ], torque: [ 0.         -5.81445992  0.        ]
contact 4, body1: R_thumb_proximal, body2: pepsi_bottle, force: [154.27413641  -0.27591384   0.22948338], torque: [ 0.90439868  1.44858855 -0.1075578 ]
contact 5, body1: R_thumb_intermediate, body2: pepsi_bottle, force: [1.95476310e+03 8.04842999e-01 2.24064008e+00], torque: [ 0.6541473  -8.42871855 88.62680457]
contact 6, body1: R_thumb_distal, body2: pepsi_bottle, force: [108.29118585   1.23837119   0.14450953], torque: [-0.50476415  2.271227   -3.381561  ]

[chaitanya1chawla]

Reducing the model.opt.impratio ratio below 1 helped reduce frictional forces, thus making it easier to drop the bottle