Implementation:Google deepmind Dm control Duplo

Knowledge Sources	Google_deepmind_Dm_control
Domains	Reinforcement_Learning, Robotics, Manipulation
Last Updated	2026-02-15 04:00 GMT

Overview

A Composer entity representing a 2x4 Duplo brick with physically realistic stud-based interlocking mechanics, configurable separation forces, and per-episode stud size randomization.

Description

The Duplo class extends composer.Entity and loads its MJCF model from an external XML file (duplo2x4.xml). The brick features a 2x4 grid of studs on top and corresponding holes on the bottom, enabling realistic snap-fit interlocking behavior between bricks. The separation force required to pull bricks apart is controlled by the stud radius, which is randomized at the start of each episode.

The _build method accepts several configuration options. The easy_align flag switches studs from cylinders to capsules, making alignment easier for robotic manipulation. The flanges flag enables bottom flanges that allow partially overlapping brick configurations at the cost of more expensive dynamics. The variation parameter (a float from 0.0 to 1.0) controls the range of stud size randomization: a value of 1.0 produces a distribution of separation forces matching empirical measurements from real Duplo bricks, while 0.0 yields a deterministic force near the mode of the empirical distribution. Stud size parameters are looked up from pre-calibrated tables indexed by the (easy_align, flanges) combination.

DuploObservables provides five sensor-based observables: position, orientation, linear velocity, angular velocity, and force, all backed by MuJoCo sensors defined in the XML model.

Usage

Use Duplo as a prop in Composer manipulation tasks involving brick stacking, assembly, or placement. Configure easy_align and flanges based on the desired difficulty and physical fidelity. Adjust variation to control the range of separation force randomization for domain randomization training.

Code Reference

Source Location

Repository: Google_deepmind_Dm_control
File: dm_control/entities/props/duplo/__init__.py
Lines: 1-170

Signature

class Duplo(composer.Entity):
    def _build(self, easy_align=False, flanges=True, variation=0.0,
               color=(1., 0., 0.)):
        ...
    def initialize_episode_mjcf(self, random_state):
        ...
    @property
    def studs(self):
        ...
    @property
    def holes(self):
        ...
    @property
    def mjcf_model(self):
        ...

class DuploObservables(composer.Observables, composer.FreePropObservableMixin):
    def position(self):
        ...
    def orientation(self):
        ...
    def linear_velocity(self):
        ...
    def angular_velocity(self):
        ...
    def force(self):
        ...

Import

from dm_control.entities.props.duplo import Duplo

I/O Contract

Inputs (_build)

Name	Type	Required	Description
easy_align	bool	No	If `True`, use capsule studs for easier alignment (default `False`)
flanges	bool	No	If `True`, enable bottom flanges for partial-overlap connections (default `True`)
variation	float	No	Controls stud size randomization range from 0.0 (deterministic) to 1.0 (full empirical range) (default 0.0)
color	tuple of 3 floats	No	RGB color values between 0 and 1 (default red: `(1., 0., 0.)`)

Outputs

Name	Type	Description
studs	`numpy.ndarray` (2, 4) of `mjcf.Element`	Site elements corresponding to the 8 studs
holes	`numpy.ndarray` (2, 4) of `mjcf.Element`	Site elements corresponding to the 8 holes
position observable	`numpy.ndarray`	Brick position from sensor data
orientation observable	`numpy.ndarray`	Brick orientation from sensor data
linear_velocity observable	`numpy.ndarray`	Brick linear velocity from sensor data
angular_velocity observable	`numpy.ndarray`	Brick angular velocity from sensor data
force observable	`numpy.ndarray`	Force sensor data on the brick

Stud Size Parameters

The stud radius is sampled uniformly between computed lower and upper bounds based on the configuration:

easy_align	flanges	minimum	lower_quartile	maximum
False	False	0.004685	0.004781	0.004898
False	True	0.004609	0.004647	0.004716
True	False	0.004754	0.004844	0.004953
True	True	0.004695	0.004717	0.004765

The variation parameter interpolates between the lower quartile (at 0.0) and the full min/max range (at 1.0).

Usage Examples

from dm_control.entities.props.duplo import Duplo

# Create a red Duplo brick with default settings
brick = Duplo()

# Create a blue brick with easier alignment and full force variation
easy_brick = Duplo(easy_align=True, flanges=True, variation=1.0,
                   color=(0., 0., 1.))

# Access studs and holes
print(brick.studs.shape)  # (2, 4)
print(brick.holes.shape)  # (2, 4)

# The stud radius is randomized at episode start
# (called automatically by the Composer environment)
import numpy as np
rng = np.random.RandomState(42)
brick.initialize_episode_mjcf(rng)

Related Pages

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment