Restitution in rigid collisions

[lankoestee]

Hi, thank you for the great work on Genesis!

I have a question regarding rigid body collisions. Specifically, how can we control the restitution (bounciness) between two rigid bodies — for example, when a rigid sphere drops and collides with a table?

I noticed that in the material definition for rigid bodies, there is a coup_restitution parameter, which seems to govern the restitution used during solver coupling. However, from my experiments, this parameter doesn't appear to influence the bouncing behavior of rigid bodies colliding with each other.

Admittedly, I can use gs.materials.MPM.Elastic to treat the sphere as a deformable elastic object and adjust the bounce using Young’s modulus and Poisson’s ratio. But I'm specifically interested in controlling bounce behavior purely within rigid body dynamics.

In comparison, the Kubric (based on PyBullet) allows setting restitution at the rigid body level, which directly affects bounce. See this line in Kubric’s object definition.

Here’s a minimal example. Even when I set all coup_restitution values to 1.0, the sphere does not bounce on the table:

import genesis as gs
import numpy as np


def main():
    gs.init(backend=gs.gpu)

    scene = gs.Scene(
        sim_options=gs.options.SimOptions(
            dt=5e-3,
            substeps=16,
        ),
        vis_options=gs.options.VisOptions(
            show_world_frame=False,
            plane_reflection=True,
            ambient_light=(0.5, 0.5, 0.5),
        ),
        show_viewer=False,
    )

    # All objects are set as rigid bodies with coup_restituion=1
    general_rigid = gs.materials.Rigid(
	    coup_friction=0.1, 
	    coup_restitution=1.0,
	)

    plane = scene.add_entity(
        material=general_rigid,
        morph=gs.morphs.Plane(),
    )

    # Put a fixed table underneath
    table = scene.add_entity(
        morph=gs.morphs.Mesh(
            file="assets/table/meshes/CoffeeTable.obj",
            scale=0.02,
            fixed=True,
        ),
        material=general_rigid,
        surface=gs.surfaces.Rough(),
    )

    # Put a ball on it and let it fall freely
    sphere = scene.add_entity(
        material=general_rigid,
        morph=gs.morphs.Sphere(
            pos=(0.0, 0.0, 1.8),
            radius=0.1,
        ),
    )

    cam = scene.add_camera(
        res=(720, 480),
        fov=45,
        GUI=False,
    )

    # Build the scene
    scene.build()

    # Added a little lens rotation
    horizon = 500
    center = np.array([0.0, 0.0, 1.0])
    radius = 3.0
    height = 2.5
    start_angle = -np.pi / 6
    end_angle = np.pi / 6

    cam.start_recording()

    for i in range(horizon):
        scene.step()

	# Camera operation
	t = i / horizon
	angle = start_angle + t * (end_angle - start_angle)
	x = center[0] + radius * np.cos(angle)
	y = center[1] + radius * np.sin(angle)
        z = height
        cam.set_pose(
            pos=(x, y, z),
	    lookat=(center[0], center[1], center[2]),
        )
        cam.render()

    cam.stop_recording(save_to_filename="video.mp4", fps=video_fps)


if __name__ == "__main__":
    main()

Is there a way to achieve this kind of rigid body bounce within Genesis? Or is restitution only meaningful when using deformable objects?

Any guidance would be greatly appreciated!

video.mp4

[Kashu7100]

@duburcqa It would be also nice if we can expose the restitution for domain randomization in RL.

[lankoestee]

Thanks, this should create a lot of convenience for rigid collision simulations without having to involve other complex solvers.

[duburcqa]

Actually, that is not how it works for Rigid Body Dynamics. The contact dynamics relies on the impedance model that was introduced by Mujoco. here is the official documentation. The behaviour of this model is specified by the vector parameters solref and solimp in Mujoco, which corresponds to sol_params in Genesis (concatenation of solref and solimp in this exact order). It stands for contact solver reference , impedance, and parameters respectively.

[gleitn96]

@duburcqa thanks for the explanation — is it possible to set sol_params individually for each rigid body (or per object-pair contact), and if so, where exactly should this be done?

[lankoestee]

@duburcqa Thank you for correcting my misunderstanding of Rigid Body Dynamics.

I noticed that during collision computation, the solver uses the average of the two objects’ sol_params, as seen here:

Genesis/genesis/engine/solvers/rigid/collider_decomp.py

Line 1163 in a0eeddd

self.contact_data[i_col, i_b].sol_params = 0.5 * (ga_info.sol_params + gb_info.sol_params)

In addition, the RigidSolver provides a method to set global sol_params:

Genesis/genesis/engine/solvers/rigid/rigid_solver_decomp.py

Lines 4027 to 4039 in a0eeddd

	def set_global_sol_params(self, sol_params):
	"""
	Solver parameters (timeconst, dampratio, dmin, dmax, width, mid, power).
	Reference: https://mujoco.readthedocs.io/en/latest/modeling.html#solver-parameters
	"""
	assert len(sol_params) == 7
	# Make sure that the constraints parameters are valid
	sol_params = _sanitize_sol_params(
	sol_params, self._sol_constraint_min_resolve_time, self._sol_constraint_resolve_time
	)
	self._kernel_set_global_sol_params(sol_params)

By default, the solver parameters for rigid bodies are initialized via the following method:

Genesis/genesis/utils/geom.py

Lines 1146 to 1153 in a0eeddd

	def default_solver_params(n=6):
	"""
	Default solver parameters (timeconst, dampratio, dmin, dmax, width, mid, power). Reference: https://mujoco.readthedocs.io/en/latest/modeling.html#solver-parameters
	Note that timeconst here will not be used in the current workflow. Instead, it will be computed using 2 * substep_dt.
	"""
	solver_params = np.array([0.0, 1.0e00, 9.0e-01, 9.5e-01, 1.0e-03, 5.0e-01, 2.0e00])
	return np.repeat(solver_params[None], n, axis=0)

---

My question is whether Genesis provides a higher-level API that supports either of the following functionalities:

1. Detect each individual collision and assign a specific sol_params to it.
2. Allow assigning custom sol_params to each rigid body during entity initialization, and compute the solver parameters at collision time by averaging the two bodies’ parameters.

I'm planning to use Genesis for rigid body simulation involving some predefined basic shapes such as cubes and spheres. If the current implementation supports either of the functionalities above, could you kindly point me to how it's done? If not, would it be possible to consider adding such features in a future release? Alternatively, would you consider updating the collision section of the user guide to cover this?

[lankoestee]

@gleitn96 I found that using MuJoCo xml for file description can achieve this task. According to the concept given in Solver Parameter, we can directly set solimp and solref in the model definition xml file, which will be weighted averaged during calculation. The following MJCF file generated by GPT can realize the function of changing the restitution coefficient.

<mujoco model="ball_drop">
  <compiler angle="degree" coordinate="local"/>
  <option timestep="0.002" gravity="0 0 -9.81" iterations="50" integrator="RK4"/>

  <!-- Default settings for all geoms -->
  <default>
    <geom solref="0.005 0.3" friction="0.5 0.1 0.0" density="1000"/>
  </default>
  <asset>
    <texture name="grid" type="2d" builtin="checker" width="512" height="512"
             rgb1="0.2 0.3 0.4" rgb2="0.1 0.2 0.3"/>
    <material name="mat_ground" texture="grid" texrepeat="4 4" reflectance="0.2"/>
  </asset>

  <worldbody>

    <!-- Static ground plane -->
    <geom name="floor" type="plane" pos="0 0 0" size="5 5 0.1" material="mat_ground" rgba="0.8 0.8 0.8 1" solref="0.005 0.3"/>
    
    <!-- Falling ball -->
    <body name="ball_0" pos="0.2 0 1">
      <freejoint/>
      <geom name="ball_geom_0" type="sphere" size="0.04" rgba="1 0 0 1" solref="0.005 0.9"/>
    </body>
    <body name="ball_1" pos="0.1 0 1">
      <freejoint/>
      <geom name="ball_geom_1" type="sphere" size="0.04" rgba="1 0 0 1" solref="0.005 0.09"/>
    </body>
    <body name="ball_2" pos="0 0 1">
      <freejoint/>
      <geom name="ball_geom_2" type="sphere" size="0.04" rgba="1 0 0 1" solref="0.005 0.009"/>
    </body>
    <body name="ball_3" pos="-0.1 0 1">
      <freejoint/>
      <geom name="ball_geom_3" type="sphere" size="0.04" rgba="1 0 0 1" solref="0.005 0.0009"/>
    </body>
    <body name="ball_4" pos="-0.2 0 1">
      <freejoint/>
      <geom name="ball_geom_4" type="sphere" size="0.04" rgba="1 0 0 1" solref="0.005 0.00009"/>
    </body>

  </worldbody>
  <actuator/>
</mujoco>

video.mp4

[duburcqa]

Nice ! I should add a unit test to reproduce this!

[duburcqa]

I added the low-level API to set the solver parameters independently for each geometry in this PR#1131:
def set_sol_params(self, sol_params, geoms_idx=None, envs_idx=None, *, joints_idx=None, eqs_idx=None, unsafe=False):

For now it is only available at rigid solver-level, but it should be exposed at joint, geom, and equality-levels soon!

[lankoestee]

Thank you for your PR! Looking forward to the release of relevant high-level APIs asap!

[duburcqa]

Here it is: #1173

[duburcqa]

I'm considering adding an even simpler API called set_restitution. Do you think it would be more convenient that sol_params?

[lankoestee]

Thank you very much! It would be better if the set_restitution API could be added. sol_params is still difficult for developers who don't fully understand mujoco.

[mi-robotics]

would be great if we could batch this as well