search

google-deepmind / mujoco

Multi-Joint dynamics with Contact. A general purpose physics simulator.
https://mujoco.org
Apache License 2.0
8.21k stars 820 forks source link

Simulated air bearings partially "sink" in the floor after I include a ball joint #1316

Closed TenguBoss closed 10 months ago

TenguBoss commented 10 months ago

How to ask for help

Aloha Colleagues,

I'm an engineer and I use MuJoCo to design and analyze a robot that slides on air bearings with (very) little friction on a (very) flat floor. I am using MiJoCo 3.1.1 on WSL2 Ubuntu Jammy.

The simplest way to explain an air bearing is an "inverted" air hokey table where the air is pumped through the puck towards the table. For detailed explanations I recommend the website of New Way Air Bearings.

I am new to MuJoCo and the only way I know how to simulate an air bearing is through frictionless contact with the floor. 

<body name="AB post -X+Y" pos=" 0.11 0.00 -0.02">
  <geom class="air bearing post"/>
  <!-- ball at bottom of post -->
  <body name="AB ball -X+Y" pos=" 0.00 0.00 -0.02">
    <geom class="air bearing ball"/>
    <!-- spherical joint between ball and pad -->
    <joint type="ball" damping="1e-3"/>
    <body name="AB pad -X+Y" pos="0 0 -0.01">
      <geom class="air bearing pad" condim="1"/>
    </body>
  </body> 
</body>

where the classes are just visualization primitives, cylinders for the post and pad and sphere for the ball. The condim="1" property of the pad geometry is matched by a condim="1" for the floor geometry.

Notes

  1. Because I am interested in the static and dynamic and balancing of the robot I cannot use slide and hinge joints at the pads. They do not allow for the robot to tip over.
  2. I have noticed an unexpected behavior. If I include a spherical joint between the post and pad, the pads partially sink in the floor and the robot has a (slight) tilt dependent on the location of the center of mass. The depth of the sink is proportional to the contact force, as illustrated in the screenshots. If I comment out the joint, the pads float on the floor, as expected.

As far as I can tell, the dynamics of robot with partially sunk pads are realistic except the tilt due to the pads thinking. I have tried to understand what is happening with the simulation and I promise I have spent a bit of time on the MuJoCo documentation to no avail.

Can you please help me understand what happens with the simulation? I think I'm missing something fundamental in setting up the spherical joints and the frictionless contact between the pads and the floor.

Thank you.

BU

Air_bearings_with_spherical_joint Air_bearings_without_spherical_joint

T0 + 1 hr update

I had a bit of time to interact with the model and have noticed that in the version with ball joints angular momenta are not conserved, within the limits of the RK4 integrator, of course. I believe this is due to the pads sinking in the floor and contributing some sort of friction or other type of energy loss mechanism. I still cannot explain it.

TenguBoss commented 10 months ago

T0 + 3 hr update (includes a lunch break)

I believe I fixed the problem of the air bearing pads sinking in the floor by adding an armature with a tiny inertia to the spherical joint

<joint name="AB ball -X+Y" type="ball" damping="1e-5" armature="1e-4"/>

However, I am leaving the topic open to get feedback on the solution.

Regards and a good weekend.

BU

yuvaltassa commented 10 months ago

a. Try using the elliptic cone option, these contacts are often better behaved. b. I just added the formulas for penetration at rest, you might want to tweak your contact parameters.