How to dynamically change body_mass #453

XinyuWang2 · 2022-08-23T21:25:36Z

XinyuWang2
Aug 23, 2022

Hi, I need to keep changing some objects' mass in the run time.
I find the following variable may be relevant:
physics.model.body_mass
physics.model.body_inertia

But when I changed it, I find the behavior of the object doesn't change at all (in some way).
Here is a minimal code that can reproduce the problem.
In the model(xml), one box is pretty heavy, and another box is trying to push it.
Even after I change the body mass to 0, it is still unable to be pushed.
But in some other way, the mass is actually changed! Because if you set the initial pos of box1 to be pos="0 0 0", it will keep flying upward.

import dm_control
import dm_control.mujoco
from dm_control import mjcf
import dm_env
import numpy as np

MODEL = '''
<mujoco model="arena">
    <visual>
        <headlight ambient="0.3 0.3 0.3" diffuse="0.6 0.6 0.6" specular="0 0 0"/>
        <map znear="0.01"/>
    </visual>

    <worldbody>
        <geom name="floor" friction="1.0" type="plane" size=".7 .7 .1" rgba="0.7 0.7 0.7 1"/>
        <body name="box1" pos="0 0 0.1">
            <freejoint/>
            <!-- <geom type="box" size="0.1 0.1 0.1" rgba="1 0 0 0.5" mass="0.01"/> -->
            <geom type="box" size="0.1 0.1 0.1" rgba="1 0 0 0.5" mass="100000"/>
        </body>

        <body name="box2" pos="0 -0.4 0.1">
            <joint name="y" type="slide" axis="0 1 0"/>
            <joint name="z" type="slide" axis="0 0 1"/>
            <geom type="box" size="0.1 0.1 0.1" rgba="0 1 0 0.5" mass="0.1"/>
        </body>

    </worldbody>

    <actuator>
        <motor name="y" joint="y" ctrllimited="true" ctrlrange="-10 10"/>
    </actuator>

</mujoco>
'''


class Env(dm_env.Environment):
    def __init__(self) -> None:
        model = mjcf.from_xml_string(MODEL)
        self.physics = mjcf.Physics.from_mjcf_model(model)

    def step(self, action) -> dm_env.TimeStep:
        self.physics.data.ctrl[:] = action[:]
        self.physics.model.body_mass[1] = 0.0
        # self.physics.model.body_inertia[1] = 0.0
        # print(self.physics.model.body_mass)

        self.physics.step()
        return dm_env.TimeStep(step_type=None,
                                reward=None,
                                discount=None,
                                observation=None)

    def reset(self) -> dm_env.TimeStep:
        return dm_env.TimeStep(step_type=dm_env.StepType.FIRST,
                               reward=None,
                               discount=None,
                               observation=None)
    def action_spec(self):
        action_spec = dm_control.mujoco.action_spec(self.physics)
        return dm_env.specs.BoundedArray(shape=action_spec.shape,
                                        dtype=action_spec.dtype,
                                        minimum=action_spec.minimum,
                                        maximum=action_spec.maximum)
    def observation_spec(self):
        pass

class Policy:

    def __init__(self, env):
        self.action_spec = env.action_spec()

    def __call__(self, time_step):
        action = np.ones(self.action_spec.shape)
        return action

env = Env()
policy = Policy(env)
from dm_control import viewer
viewer.launch(env, policy=policy)

yuvaltassa · 2022-08-23T21:43:34Z

yuvaltassa
Aug 23, 2022
Maintainer

Inertia is a 3-vector.
You need to scale the mass and inertia in the same way to get correct physics, so use scaling, not setting:

model.body_mass[i] *= c
model.body_inertia[i, :] *= c

8 replies

fukangl Jul 18, 2024

@yuvaltassa When I scale the body mass and inertia using:

model.body_mass[i] *= c
model.body_inertia[i, :] *= c

There are errors such as
WARNING:absl:Inertia matrix is too close to singular at DOF 47. Check model. Time = 0.0000.
and
WARNING:absl:Nan, Inf or huge value in QACC at DOF 45. The simulation is unstable. Time = 470.7715.

The scaling factor is randomly chosen between [0.5,1.5], and even when I use a smaller range [0.8,1.2], this issue still occurs. Do you have any suggestions on how to resolve this problem?

yuvaltassa Jul 19, 2024
Maintainer

I recommend changing the values in the mjSpec and recompiling.

fukangl Jul 22, 2024

@yuvaltassa Thanks a lot! I use the following code to compile the mjSpec. However, no matter what value I use for self.spec.settotalmass, the performance of a pre-trained policy (without mass scaling) behaves similarly. I was wondering if this is the correct way to scale the mass and inertia. Any suggestions are appreciated!

self.spec = mujoco.MjSpec()
self.spec.from_file(path_to_xml_out)
self.model = self.spec.compile()
self.data = mujoco.MjData(self.model)
self.spec.settotalmass = 50
self.model, self.data = self.spec.recompile(self.model, self.data)

fukangl Jul 22, 2024

During the test, when I use the following code, the policy behaves reasonably. For example, the policy performs best when c=1 and worse when c is much smaller than 1

model.body_mass[i] *= c
model.body_inertia[i, :] *= c

fukangl Jul 22, 2024

It seems this problem has been solved. Just recompile the XML for each training episode, although it requires more memory resources

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

How to dynamically change body_mass #453

Uh oh!

{{title}}

Uh oh!

Uh oh!

{{editor}}'s edit

{{editor}}'s edit

Uh oh!

Replies: 1 comment 8 replies

Uh oh!

{{title}}

Uh oh!

Uh oh!

{{title}}

Uh oh!

Uh oh!

{{editor}}'s edit

{{editor}}'s edit

Uh oh!

Uh oh!

{{title}}

Uh oh!

Uh oh!

{{title}}

Uh oh!

Uh oh!

{{title}}

Uh oh!

Uh oh!

{{title}}

Uh oh!

Select a reply

Uh oh!

How to dynamically change body_mass #453

Uh oh!

Uh oh!

XinyuWang2 Aug 23, 2022

Replies: 1 comment · 8 replies

Uh oh!

yuvaltassa Aug 23, 2022 Maintainer

Uh oh!

Uh oh!

fukangl Jul 18, 2024

Uh oh!

yuvaltassa Jul 19, 2024 Maintainer

Uh oh!

fukangl Jul 22, 2024

Uh oh!

fukangl Jul 22, 2024

Uh oh!

fukangl Jul 22, 2024

XinyuWang2
Aug 23, 2022

Replies: 1 comment 8 replies

yuvaltassa
Aug 23, 2022
Maintainer

yuvaltassa Jul 19, 2024
Maintainer