How to create a custom object#

Objects, such as boxes and cans, are essential to building manipulation environments. We designed the MujocoObject interfaces to standardize and simplify the procedure for importing 3D models into the scene or procedurally generate new objects. MuJoCo defines models via the MJCF XML format. These MJCF files can either be stored as XML files on disk and loaded into simulator, or be created on-the-fly by code prior to simulation. Based on these two mechanisms of how MJCF models are created, we offer two main ways of creating your own object:

  • Define an object in an MJCF XML file;

  • Use procedural generation APIs to dynamically create an MJCF model.

The MujocoObject class#

class MujocoObject(MujocoModel):
    def __init__(...):

        # Attributes that should be filled in within the subclass
        self._name = None
        self._obj = None

        # Attributes that are auto-filled by _get_object_properties call
        self._root_body = None
        self._bodies = None
        self._joints = None
        self._actuators = None
        self._sites = None
        self._contact_geoms = None
        self._visual_geoms = None

MujocoObject is the base class of all objects. One must note that it is not a subclass of MujocoXML, but does extend from the unifying MujocoModel class from which all simulation models (including robots, grippers, etc.) should extend from. All of the attributes shown above prepended with a _ are intended to be private variables and not accessed by external objects. Instead, any of these properties can be accessed via its public version, without the _ (e.g.: to access all the object’s joints, call obj.joints instead of obj._joints). This is because all public attributes are automatically post-processed from their private counterparts and have naming prefixes appended to it.

The XML of an object is generated once during initialization via the _get_object_subtree call, after which any external object can extract a reference to this XML via the get_obj call.

    def _get_object_subtree(self):

    def get_obj(self):

Additionally, objects are usually placed relatively. For example, we want to put an object on a table or place a cube on top of another. Instance methods get_bottom_offset, get_top_offset, get_horizontal_radius provide the necessary information to place objects properly.

    def get_bottom_offset(self):

    def get_top_offset(self):

    def get_horizontal_radius(self):

This allows us to do things like the following.

table_top = np.array([0, 1, 0])
bottom_offset = obj.get_bottom_offset()
pos = table_top - bottom_offset                             # pos + bottom_offset = table_top
obj_xml = obj.get_obj().set("pos", array_to_string(pos))    # Set the top-level body of this object

Creating a XMLObject#

One can use MuJoCo MJCF XML to generate an object, either as a composition of primitive geoms or imported from STL files of triangulated meshes. An example is robosuite.models.objects.xml_objects.BreadObject. Its python definition is short. Note that all MujocoXMLObject classes require both a fname and name argument, the former which specifies the filepath to the raw XML file and the latter which specifies the in-sim name of the object instantiated. The optional joints argument can also specify a custom set of joints to apply to the given object (defaults to “default”, which is a single free joint). Additionally, the type of object created can be specified via the obj_type argument, and must be one of ('collision', 'visual', or 'all'). Lastly, setting duplicate_collision_geoms makes sure that all collision geoms automatically have an associated visual geom as well. Generally, the normal use case is to define a single class corresponding to a specific XML file, as shown below:

class BreadObject(MujocoXMLObject):
    def __init__(self, name):
                         name=name, joints=[dict(type="free", damping="0.0005")],
                         obj_type="all", duplicate_collision_geoms=True)

In the corresponding XML file, a few key definitions must be present. The top-level, un-named body must contain as immediate children tags (a) the actual object bodie(s) (the top-level must be named object) and (b) three site tags named bottom_site, top_site, and horizontal_radius_site and whose pos values must be specified. The example for the BreadObject, bread.xml, is shown below:

<mujoco model="bread">
    <mesh file="meshes/bread.stl" name="bread_mesh" scale="0.8 0.8 0.8"/>
    <texture file="../textures/bread.png" type="2d" name="tex-bread" />
    <material name="bread" reflectance="0.7" texrepeat="15 15" texture="tex-bread" texuniform="true"/>
      <body name="object">
        <geom pos="0 0 0" mesh="bread_mesh" type="mesh" solimp="0.998 0.998 0.001" solref="0.001 1" density="50" friction="0.95 0.3 0.1"  material="bread" group="0" condim="4"/>
      <site rgba="0 0 0 0" size="0.005" pos="0 0 -0.045" name="bottom_site"/>
      <site rgba="0 0 0 0" size="0.005" pos="0 0 0.03" name="top_site"/>
      <site rgba="0 0 0 0" size="0.005" pos="0.03 0.03 0" name="horizontal_radius_site"/>


  • _get_object_subtree looks for the object bodie(s) as defined by all nested bodie(s) beginning with the object-named body tag.

  • bottom_site should be the bottom of the object, i.e. contact point with the surface it is placed on.

  • top_site should be the top of the object, i.e. contact point if something is placed on it.

  • horizontal_radius_site can be any point on a circle in the x-y plane that does not intersect the object. This allows us to place multiple objects without having them collide into one another.

Creating a procedurally generated object#

Procedurally generated objects have been used in several recent works to train control policies with improved robustness and generalization. Here you can programmatically generate an MJCF XML of an object from scratch using xml.etree.ElementTree, and compose an object of multiple geom primitives. The base class for this type of object is MujocoGeneratedObject. robosuite natively supports all Mujoco primitive objects with procedurally-generated PrimitiveObject classes (BoxObject, BallObject, CapsuleObject, and CylinderObject).

Additionally, robosuite supports custom, complex objects that can be defined by collections of primitive geoms (the CompositeObject class) or even other objects (the CompositeBodyObject class). The APIs for each of these classes have been standardized for ease of usage, and interested readers should consult the docstrings for each of these classes, as well as provided examples of each class (HammerObject, HingedBoxObject).

It should also be noted that all of the above classes extending from the MujocoGenereatedObject class automatically supports custom texture definitions on a per-geom level, where specific texture images can be mapped to individual geoms. The above HammerObject showcases an example applying custom textures to different geoms of the resulting object.

Placing Objects#

Object locations are initialized on every environment reset using instances of the ObjectPositionSampler class. Object samplers use the bottom_site and top_site sites of each object in order to place objects on top of other objects, and the horizontal_radius_site site in order to ensure that objects do not collide with one another. The most basic sampler is the UniformRandomSampler class - this just uses rejection sampling to place objects randomly. As an example, consider the following code snippet from the __init__ method of the Lift environment class.

self.placement_initializer = UniformRandomSampler(
    x_range=[-0.03, 0.03],
    y_range=[-0.03, 0.03],

This will sample the self.cube’s object location uniformly at random in a box of size 0.03 (x_range, y_range) with random (rotation) z-rotation (rotation_axis), and with an offset of 0.01 (z_offset) above the table surface location (reference_pos). The sampler will also make sure that the entire object boundary falls within the sampling box size (ensure_object_boundary_in_range) and does not collide with any placed objects (ensure_valid_placement).

Another common sampler is the SequentialCompositeSampler, which is useful for composing multiple arbitrary placement samplers together. As an example, consider the following code snippet from the __init__ method of the NutAssembly environment class.

# Establish named references to each nut object
nut_names = ("SquareNut", "RoundNut")

# Initialize the top-level sampler
self.placement_initializer = SequentialCompositeSampler(name="ObjectSampler")

# Create individual samplers per nut
for nut_name, default_y_range in zip(nut_names, ([0.11, 0.225], [-0.225, -0.11])):
            x_range=[-0.115, -0.11],
# No objects have been assigned to any samplers yet, so we do that now
for i, (nut_cls, nut_name) in enumerate(zip(
        (SquareNutObject, RoundNutObject),
    nut = nut_cls(name=nut_name)
    self.placement_initializer.add_objects_to_sampler(sampler_name=f"{nut_name}Sampler", mujoco_objects=nut)

The code snippet above results in two UniformRandomSampler instances being used to place the nuts onto the table surface - one for each type of nut. Notice this also allows the nuts to be initialized in separate regions of the table, and with arbitrary sampling settings. The SequentialCompositeSampler makes it easy to compose multiple placement initializers together and assign objects to each sub-sampler in a modular way.