Microduck RL MJCF Robot Models: Complete Guide to All 8 XML Configurations

Microduck RL defines eight distinct MJCF robot models—including walking, ground-contact, roller-skate, and backlash variants—each declared in src/mjlab_microduck/robot/microduck_constants.py and loaded via dedicated helper functions.

The Microduck RL framework from pollen-robotics/microduck_rl uses MuJoCo MJCF XML files to describe every physical robot configuration used across its reinforcement learning environments. This article breaks down each model's purpose, file location, and how to load it in your code.

---

The Central constants.py File

All MJCF robot models are declared in src/mjlab_microduck/robot/microduck_constants.py. According to the Microduck RL source code, this file contains:

• Path constants for each XML file (e.g., MICRODUCK_WALK_XML, MICRODUCK_GROUNDCONTACT_XML)
• Helper functions that return compiled mujoco.MjSpec objects
• Version-specific variants for tasks requiring different collision fidelity or mechanical realism

Lines 16-36 define the complete mapping between model names and their filesystem locations.

---

Base Walking Models

Walk Robot (robot_walk.xml)

The baseline bipedal robot used for standard walking and velocity-tracking tasks.

Attribute	Value
Path constant	MICRODUCK_WALK_XML (lines 16-17)
XML path	src/mjlab_microduck/robot/microduck/robot_walk.xml
Load function	get_walk_spec()

from mjlab_microduck.robot.microduck_constants import get_walk_spec

# Load the standard walking robot spec

walk_spec = get_walk_spec()

Ground-Contact Robot (robot_groundcontact.xml)

A curated collision subset for precise foot-ground interaction. Includes collision geoms for soles, legs, trunk, head, jaw, battery, and hips.

Attribute	Value
Path constant	MICRODUCK_GROUNDCONTACT_XML (lines 18-20)
XML path	src/mjlab_microduck/robot/microduck/robot_groundcontact.xml
Load functions	get_standup_spec(), get_ground_pick_spec()

Used in standup and ground manipulation tasks where contact fidelity matters.

---

Advanced Collision Models

All-Collisions Robot (robot_allcollisions.xml)

Full collision coverage with 70 geoms and 37 meshes. Every robot part carries collision geometry.

Attribute	Value
Path constant	MICRODUCK_ALLCOLLISIONS_XML (lines 22-25)
XML path	src/mjlab_microduck/robot/microduck/robot_allcollisions.xml
Load function	get_allcollisions_spec()

Currently unused in active environments but exported for future tasks requiring maximum physical accuracy.

---

Roller-Skate Variants

Ground-Contact Robot with Rollers (robot_groundcontact_rollers.xml)

Extends the walk model with passive wheel hinges (passive_*wheel) for roller-skate locomotion tasks.

Attribute	Value
Path constant	MICRODUCK_GROUNDCONTACT_ROLLERS_XML (lines 28-30)
XML path	src/mjlab_microduck/robot/microduck/robot_groundcontact_rollers.xml
Load function	get_walk_rollers_spec()

---

Backlash Models for Real-World Fidelity

Microduck RL includes backlash variants that add passive hinges with ±1° of play to every servo joint, matching real-world encoder behavior.

Backlash Ground-Contact Robot (robot_groundcontact_backlash.xml)

Attribute	Value
Path constant	MICRODUCK_GROUNDCONTACT_BACKLASH_XML (lines 30-34)
Load function	get_backlash_spec()

Backlash Walk Robot (robot_walk_backlash.xml)

Attribute	Value
Path constant	MICRODUCK_WALK_BACKLASH_XML (lines 33-35)
Load function	get_walk_backlash_spec()

Backlash Rollers Robot (robot_groundcontact_rollers_backlash.xml)

Combines roller-skate mechanics with backlash simulation—the most complex configuration.

Attribute	Value
Path constant	MICRODUCK_GROUNDCONTACT_ROLLERS_BACKLASH_XML (lines 35-36)
Load function	get_rollers_backlash_spec()

from mjlab_microduck.robot.microduck_constants import (
    get_rollers_backlash_spec,
    MICRODUCK_GROUNDCONTACT_ROLLERS_BACKLASH_XML,
)

# Most complex model: rollers + backlash

spec = get_rollers_backlash_spec()
print("XML source:", MICRODUCK_GROUNDCONTACT_ROLLERS_BACKLASH_XML)

---

Auxiliary Object Models

Ball Object (ball.xml)

A 70 mm, 15 g ball used exclusively in the BallKick task.

Attribute	Value
Path constant	MICRODUCK_BALL_XML (lines 26-28)
XML path	src/mjlab_microduck/robot/microduck/ball.xml
Load function	get_ball_spec()

from mjlab_microduck.robot.microduck_constants import get_ball_spec

ball_spec = get_ball_spec()  # For ball-kicking environments

---

Loading and Inspecting MJCF Models

Inspect Any XML File Path

from mjlab_microduck.robot.microduck_constants import MICRODUCK_WALK_XML

print("Walk model XML:", MICRODUCK_WALK_XML)

# Output: src/mjlab_microduck/robot/microduck/robot_walk.xml

Complete Model Selection Pattern

Environment configuration files in src/mjlab_microduck/tasks/ (e.g., microduck_velocity_env_cfg.py, microduck_roller_crouch_env_cfg.py) import these specs to instantiate the appropriate robot model for each training task.

---

Related Source Files

File Path	Purpose
src/mjlab_microduck/robot/microduck_constants.py	Central MJCF path definitions and loader functions
src/mjlab_microduck/tasks/*_env_cfg.py	Environment configs that select robot models
src/mjlab_microduck/robot/microduck/*.xml	Actual MJCF XML files
src/mjlab_microduck/actuator/*	BAM and backlash encoder configurations

---

Summary

• Eight MJCF robot models are defined in microduck_constants.py (lines 16-36)
• Three complexity tiers: base walking, ground-contact (curated collisions), and all-collisions (full geometry)
• Three mechanical variants: standard, roller-skate (passive wheels), and backlash (±1° joint play)
• One auxiliary object: 70 mm ball for kicking tasks
• Consistent API: each model has a path constant (MICRODUCK_*_XML) and loader function (get_*_spec())

---

Frequently Asked Questions

How do I choose between walk and ground-contact models?

robot_walk.xml is lighter and faster for basic locomotion tasks. robot_groundcontact.xml adds precise sole and leg collisions when foot-ground interaction accuracy matters—as in the standup and ground-pick environments.

What does the backlash model simulate?

Real servo encoders have mechanical play. The backlash models add passive hinges with ±1° of free movement to every joint, making simulation trajectories match physical robot behavior more closely. Use these when validating policies on hardware.

Can I use the all-collisions model in training?

The all-collisions robot (robot_allcollisions.xml with 70 geoms) is currently unused in active environments. It exists for future tasks where full-body collision detection outweighs the computational cost.

Where are these models actually instantiated?

Environment configuration files in src/mjlab_microduck/tasks/ import specs from microduck_constants.py. For example, roller-skate tasks call get_walk_rollers_spec() or get_rollers_backlash_spec() depending on whether real-world fidelity is required.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}