google-deepmind
Abnormal slip
Intro
Hi!
I am a student and I'm using Mujoco for rigid body simulation and reinforcement learning.
My setup
Both 3.3.0 and the Jupyter notebook in https://mujoco.readthedocs.io/en/stable/python.html.
What's happening? What did you expect?
I found that when I use a large-magnitude, high-frequency driving force to actuate a joint, noticeable and abnormal slip occurs. I have read the explanation about slip in the tutorial and understand that MuJoCo does not strictly prevent it from happening. However, I am not sure whether the situation I am facing matches the description in that section, because changing solref, solimp, and impratio does not seem to help.
I would like to know how I can ensure simulation accuracy under this kind of control force setup.
Steps for reproduction
Just need to run the following Python script.
Minimal model for reproduction
<mujoco model="Test">
<option timestep="0.0001">
<flag eulerdamp="disable" contact="enable" gravity="enable"/>
</option>
<asset>
<texture type="skybox" builtin="gradient" rgb1=".3 .5 .7" rgb2="0 0 0" width="32" height="512"/>
<texture name="body" type="cube" builtin="flat" mark="cross" width="128" height="128" rgb1="0.8 0.6 0.4" rgb2="0.8 0.6 0.4" markrgb="1 1 1" random="0.01"/>
<material name="body" texture="body" texuniform="true" rgba="0.8 0.6 .4 1"/>
<texture name="grid" type="2d" builtin="checker" width="512" height="512" rgb1=".1 .2 .3" rgb2=".2 .3 .4"/>
<material name="grid" texture="grid" texrepeat="1 1" texuniform="true" reflectance=".2"/>
</asset>
<worldbody>
<light name="light" pos="-.2 0 1"/>
<geom name="ground" type="plane" pos="0 0 0" size="0 0 10" material="grid"
zaxis="0 0 1" friction="1.0"/>
<light pos="0 0 1"/>
<light pos="1 0 1"/>
<light pos="2 0 1"/>
<light pos="3 0 1"/>
<light pos="4 0 1"/>
<light pos="5 0 1"/>
<camera name="cam" pos="0.0 -1.0 1.0" xyaxes="1 0 0 0 1 1" fovy="120" />
<body name="b1" pos="0 0 0.05">
<joint name="j0" type="free"/>
<geom name="g1" type="capsule" size="0.05" fromto="-0.05 0 0 0.05 0 0" friction="1.0"/>
<body name="b2" pos="0.2 0 0">
<joint name="j1" type="hinge" axis="0 0 1" pos="-0.1 0 0" range="-70 70"/>
<geom name="g2" type="capsule" size="0.05" fromto="-0.05 0 0 0.05 0 0" friction="1.0"/>
</body>
</body>
</worldbody>
<contact>
<pair geom1="g1" geom2="ground" friction="0.01 0.01"/>
<pair geom1="g2" geom2="ground" friction="0.01 0.01"/>
<exclude body1="b1" body2="b2"/>
</contact>
<actuator>
<motor name="act1" gear="10" joint="j1"/>
</actuator>
</mujoco>
Code required for reproduction
import os
import numpy as np
import mediapy as media
import mujoco
if __name__ == "__main__":
import os
mj_model = mujoco.MjModel.from_xml_string(xml)
mj_data = mujoco.MjData(mj_model)
renderer = mujoco.Renderer(mj_model)
scene_option = mujoco.MjvOption()
scene_option.flags[mujoco.mjtVisFlag.mjVIS_JOINT] = True
duration = 3 # (seconds)
framerate = 24 # (Hz)
frames = []
mujoco.mj_resetData(mj_model, mj_data)
n_step = 0
i = 0
xfrc = np.zeros_like(mj_data.xfrc_applied)
while mj_data.time < duration:
ctrl = np.array([(-1) ** (i // 1000)])
i += 1
mj_data.ctrl = ctrl
mujoco.mj_step(mj_model, mj_data)
if len(frames) < mj_data.time * framerate:
renderer.update_scene(mj_data, scene_option=scene_option)
pixels = renderer.render()
frames.append(pixels)
media.show_video(frames, fps=framerate)
Confirmations
- [x] I searched the latest documentation thoroughly before posting.
- [x] I searched previous Issues and Discussions, I am certain this has not been raised before.
I ran your code (adjusted to make a 10x slow motion video).
https://github.com/user-attachments/assets/52423393-49a3-440a-9b2f-79ee4b192dad
The slip described in the documentation is about situations where objects are under a force that is within the friction cone so shouldn't be moving, but they do move a little bit, so different from what you have here.
Can you explain what you don't like about this video? Are you talking about the slow movement left?
Yes, I'm talking about the slow movement left exactly. I wonder why it happens and how I can avoid this weird phenomenon.
This looks like it could be caused by the non-isotropic nature of pyramidal friction. Can you try elliptic?
I tried this solution. Unluckily this phenomenon still exists (the following video 1). Also I tried closing frictional force during the simulation, and found that the abnormal slip still exists but performs a little different (shown in the following video 2).
https://github.com/user-attachments/assets/881a3fde-fc93-41f1-8671-20ffab0b61bb
https://github.com/user-attachments/assets/004729d9-8b2f-44d1-a592-3633c87f2152
Can you please check if one of the above examples is different with smaller timestep? I realize you're already using 1e-4, but it would be informative if you tried 1e-6 and got the same/different results
After choosing 1e-7 as dt with the object oscillating with the same frequency, it looks like this phenomenon still exists, but is slightly less pronounced. But if I set the friction coefficients as 0, then it becomes correct.
https://github.com/user-attachments/assets/cd485dc7-6e02-4c09-8f81-10d761c9562a
https://github.com/user-attachments/assets/bc20a0f7-5e37-4b97-8871-a4ae110b71ff
Also, by using dt = 1e-6 I obtained similar results with these two videos.