Why can contact -dist be < cylinder radius even when the box visually fully covers the cylinder cross-section? (box vs cylinder, MuJoCo 3.3.4) #2797

kankanzheli · 2025-08-10T12:33:31Z

kankanzheli
Aug 10, 2025

Intro

Hi!

I am a doctor student at Waseda unv, I use MuJoCo for my research on reinforcement learning.I need to simulate the collision between plants and robotic arms.

My setup

MuJoCo: 3.3.4 (Python package and C library)

OS: Ubuntu (GLFW rendering)

GL backend: MUJOCO_GL=glfw

My question

What I’m doing
I’m studying contact depth during a simple collision between a box and a slender cylinder (axis along z, i.e., circular cross-section lies in xy). Both bodies move along x slide joints toward each other. Each simulation step I print the maximal contact depth as pen = max(0.0, -con.dist) and also a “radial depth” I compute myself for the cylinder (project contact point to the cylinder axis and use radius - radial_distance).

Observation (unexpected)

As the bodies start to overlap slightly, -con.dist increases as expected (always < radius).

Once the cylinder center enters the box and the scene evolves toward the box visually covering the whole circular cross-section, the printed -con.dist starts to decrease.

Even when they look completely overlapped in that cross-section view, the reported -con.dist remains below the cylinder radius.

I understand -con.dist is the minimum translation distance (MTD) from GJK/EPA in 3D. My hypothesis is that when the active feature changes (face→edge→vertex, or cylinder side→cap vs box corner/edge), the shortest separation direction becomes oblique (often toward a nearby edge/corner, possibly with an axial component), so the 3D MTD can indeed be smaller than the cross-sectional (radial) overlap. I want to confirm whether this interpretation is correct for cylinder–box contacts in MuJoCo.

Questions

Is -con.dist exactly the depth used by MuJoCo’s soft contact model to compute the normal force, or is there any other penetration-like quantity internally used together with solref/solimp?

For cylinders/capsules, is there a recommended way to obtain a monotonic “radial overlap depth” consistent with cross-section intuition (e.g., project the contact point to the cylinder axis and use radius - distance_to_axis)?

If MuJoCo already exposes such a metric, which field is it?

Are there recommended settings to reduce feature switching (so MTD aligns more with the box face normal), for example by using certain solimp/solref/margin choices or by modeling choices (capsule vs cylinder vs mesh)?

Is there a way to debug which features are active (face/edge/vertex) for the current contact in MuJoCo (any debug flags or fields to inspect)?

What I expected vs saw

I expected that once the cross-section looks fully overlapped, the depth would at least be around the cylinder radius in that view.

What I see is that -con.dist peaks and then decreases as the situation evolves toward full visual coverage of the circular cross-section, and it stays below the radius even when visually fully overlapped.

Could you please confirm:

whether this is the intended behavior of -con.dist (3D MTD) under feature changes for cylinder–box; and

if there’s a recommended MuJoCo-native metric or workflow to obtain a cross-section-consistent (monotonic) depth for cylinders/capsules?

Thank you!

Minimal model and/or code that explain my question

Code:

# box_cylinder_viewer.py
import os, time, numpy as np
os.environ.setdefault("MUJOCO_GL", "glfw")

import mujoco
import mujoco.viewer
from textwrap import dedent

XML = dedent("""
<mujoco model="box_cylinder_pen_demo">
  <option timestep="0.001" gravity="0 0 0" integrator="implicit"/>
  <default>
    <!-- soften + zero friction just to observe geometry -->
    <geom condim="3" solref="0.3 1" solimp="0.95 0.99 0.02 0 2" friction="0 0 0"/>
  </default>

  <worldbody>
    <camera name="top" mode="fixed" pos="0 0 0.25" euler="-90 0 0"/>

    <!-- box slides along +x from the left -->
    <body name="box_body" pos="-0.06 0 0">
      <joint name="box_slide" type="slide" axis="1 0 0"/>
      <geom  name="box" type="box" size="0.015 0.015 0.015"/>
    </body>

    <!-- slender cylinder, axis along z (cross-section in xy), slides along -x from the right) -->
    <!-- cylinder size = [radius, half_length] -->
    <body name="cyl_body" pos="0.06 0 0">
      <joint name="cyl_slide" type="slide" axis="1 0 0"/>
      <geom  name="cyl" type="cylinder" size="0.005 0.0125"/>
    </body>
  </worldbody>

  <actuator>
    <velocity name="v_box" joint="box_slide" kv="200" ctrlrange="-2 2"/>
    <velocity name="v_cyl" joint="cyl_slide" kv="200" ctrlrange="-2 2"/>
  </actuator>
</mujoco>
""").strip()

def radial_pen_cylinder(model, data, cyl_gid, contact_pos):
    """radial overlap depth: radius - distance(contact point to cylinder axis segment)."""
    R = data.geom_xmat[cyl_gid].reshape(3, 3)
    axis = R[:, 2]; axis = axis / (np.linalg.norm(axis) + 1e-12)
    C = data.geom_xpos[cyl_gid].copy()
    r = model.geom_size[cyl_gid, 0]; half = model.geom_size[cyl_gid, 1]
    A = C - half*axis; B = C + half*axis
    AB = B - A; AP = contact_pos - A
    ab2 = float(np.dot(AB, AB))
    t = 0.5 if ab2 < 1e-12 else float(np.clip(np.dot(AP, AB)/ab2, 0.0, 1.0))
    Q = A + t*AB
    dr = float(np.linalg.norm(contact_pos - Q))
    return max(0.0, r - dr), r

def main():
    model = mujoco.MjModel.from_xml_string(XML)
    data  = mujoco.MjData(model)

    gid_box = mujoco.mj_name2id(model, mujoco.mjtObj.mjOBJ_GEOM, "box")
    gid_cyl = mujoco.mj_name2id(model, mujoco.mjtObj.mjOBJ_GEOM, "cyl")
    aid_box = mujoco.mj_name2id(model, mujoco.mjtObj.mjOBJ_ACTUATOR, "v_box")
    aid_cyl = mujoco.mj_name2id(model, mujoco.mjtObj.mjOBJ_ACTUATOR, "v_cyl")

    v_box = +0.15; v_cyl = -0.15  # m/s

    print("time\tncon\tpen_max(MTD)\tpen_radial_max\tradius")
    with mujoco.viewer.launch_passive(model, data) as viewer:
        while viewer.is_running():
            step_start = time.time()
            data.ctrl[aid_box] = v_box
            data.ctrl[aid_cyl] = v_cyl
            mujoco.mj_step(model, data)

            pens, pens_rad = [], []
            for i in range(data.ncon):
                con = data.contact[i]
                if not ({con.geom1, con.geom2} == {gid_box, gid_cyl}):
                    continue
                pen = max(0.0, -con.dist)
                pos = np.array(con.pos, dtype=float)
                pen_r, r = radial_pen_cylinder(model, data, gid_cyl, pos)
                pens.append(pen); pens_rad.append(pen_r)

            pen_max = max(pens) if pens else 0.0
            pen_rad_max = max(pens_rad) if pens_rad else 0.0
            radius = model.geom_size[gid_cyl, 0]
            print(f"{data.time:.4f}\t{len(pens):d}\t{pen_max:.6f}\t{pen_rad_max:.6f}\t{radius:.6f}", flush=True)

            viewer.sync()
            time.sleep(max(0.0, model.opt.timestep - (time.time() - step_start)))

if __name__ == "__main__":
    main()

Confirmations

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

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Why can contact -dist be < cylinder radius even when the box visually fully covers the cylinder cross-section? (box vs cylinder, MuJoCo 3.3.4) #2797

Uh oh!

{{title}}

Uh oh!

Replies: 0 comments

Select a reply

Uh oh!

Why can contact -dist be < cylinder radius even when the box visually fully covers the cylinder cross-section? (box vs cylinder, MuJoCo 3.3.4) #2797

Uh oh!

kankanzheli Aug 10, 2025

Intro

My setup

My question

Minimal model and/or code that explain my question

Confirmations

Replies: 0 comments

kankanzheli
Aug 10, 2025