Microduck RL Locomotion Tasks: Complete Task Catalog for Bipedal Robot Training

Microduck RL provides 18+ distinct locomotion tasks spanning velocity control, recovery behaviors, manipulation, and acrobatic skills, each available in flat or rough terrain variants plus optional backlash simulation.

The pollen-robotics/microduck_rl repository implements a comprehensive task library for training a bipedal robot using reinforcement learning. All tasks are centrally registered in src/mjlab_microduck/tasks/__init__.py through MJLab's register_mjlab_task helper, which binds environment configurations, RL hyperparameters, and execution runners into trainable units.

Core Locomotion Task Families

Microduck RL organizes its locomotion tasks into functional families based on robot behavior and terrain complexity. Each family supports both flat and rough terrain variants unless noted otherwise.

Standard Velocity Control

The foundational walking tasks train the robot to track linear and angular velocity commands.

• Mjlab-Velocity-Flat-MicroDuck — Baseline velocity tracking on flat ground
• Mjlab-Velocity-Rough-MicroDuck — Velocity tracking with randomized terrain height

Both use make_microduck_velocity_env_cfg() with rough=True toggling terrain complexity. The RL configuration class is MicroduckRlCfg.

VelStand: Walk, Recovery, and Pose Control

The VelStand task family combines velocity tracking with fall recovery and explicit pose control, making it robust for real-world deployment.

• Mjlab-VelStand-Flat-MicroDuck
• Mjlab-VelStand-Rough-MicroDuck

Configuration via make_microduck_velstand_env_cfg() and MicroduckVelStandRlCfg. This is the recommended starting point for general locomotion policies.

Stand-Up Recovery

Trains the robot to recover from an inverted (back-lying) starting position.

• Mjlab-StandUp-Flat-MicroDuck
• Mjlab-StandUp-Rough-MicroDuck

Uses make_microduck_standup_env_cfg() and MicroduckStandUpRlCfg. Critical for fault-tolerant deployment.

Sit/Stand Transitions

Explicit commanded transitions between sitting and standing postures.

• Mjlab-SitStand-Flat-MicroDuck
• Mjlab-SitStand-Rough-MicroDuck

Configuration via make_microduck_sitstand_env_cfg() and MicroduckSitStandRlCfg. Useful for power-saving behaviors and human-robot interaction.

Manipulation and Interaction Tasks

Ground-Pick: Crouch-to-Grasp Behaviors

Trains a multi-phase behavior: crouch → mouth contact → stand with object.

• Mjlab-GroundPick-Flat-MicroDuck
• Mjlab-GroundPick-Rough-MicroDuck

Uses make_microduck_ground_pick_env_cfg() and MicroduckGroundPickRlCfg. The task specification includes contact detection between the robot's mouth and target objects.

Ball Kick

A flat-terrain-only task for kicking a 70 mm ball toward a target direction.

• Mjlab-BallKick-Flat-MicroDuck

Configuration via make_microduck_ball_kick_env_cfg() and MicroduckBallKickRlCfg. Adds spherical object dynamics to the simulation.

Roller-Equipped Variants

Microduck RL includes a distinct hardware variant with passive roller wheels on the feet, enabling skating gaits.

Roller Velocity Control

• Mjlab-Velocity-Flat-MicroDuck-Rollers — Standard velocity on roller hardware

Uses make_microduck_velocity_rollers_env_cfg() and MicroduckRollersRlCfg.

Roller-Swizzle: Symmetric Foot-Grounded Gait

• Mjlab-Velocity-Swizzle-MicroDuck — Symmetric velocity control with both feet grounded

Uses make_microduck_velocity_swizzle_env_cfg() and MicroduckSwizzleRlCfg. Unique among roller tasks for its bilateral constraint.

Specialized Roller Behaviors

Task ID	Description	Config Function	RL Config
Mjlab-RollerCrouch-Flat-MicroDuck	Stable crouched posture on rollers	make_microduck_roller_crouch_env_cfg()	MicroduckRollerCrouchRlCfg
Mjlab-RollerSlope-Flat-MicroDuck	Navigation on inclined terrain	make_microduck_roller_slope_env_cfg()	MicroduckRollerSlopeRlCfg
Mjlab-RollerStandUp-Flat-MicroDuck	Recovery to stand on roller feet	make_microduck_roller_standup_env_cfg()	MicroduckRollerStandUpRlCfg

Acrobatic and Specialized Locomotion Tasks

Spin: On-Spot Rotation

• Mjlab-Spin-Flat-MicroDuck — High-speed in-place rotation

Uses make_microduck_spin_env_cfg() and MicroduckSpinRlCfg. Optimizes for maximum yaw rate while maintaining balance.

Roulade: Forward Roll

• Mjlab-Roulade-Flat-MicroDuck — Complete forward roll over the head

Uses make_microduck_roulade_env_cfg() and MicroduckRouladeRlCfg. The most dynamic task in the library, requiring precise timing of limb retraction and extension.

Backlash Variants for Sim-to-Real Transfer

Every base task has a backlash-augmented counterpart that injects ±1° gear-play and encoder backlash. These variants use an identical task_id pattern with -Backlash- inserted:

Mjlab-Velocity-Flat-Backlash-MicroDuck
Mjlab-StandUp-Rough-Backlash-MicroDuck
Mjlab-VelStand-Flat-Backlash-MicroDuck

The registration loop in src/mjlab_microduck/tasks/__init__.py (lines 54-71) automatically generates these variants using the helper in backlash.py. Training with backlash improves robustness to real hardware imperfections.

Quick-Start Code Examples

List All Available Tasks

from mjlab.tasks.registry import list_tasks

for task in list_tasks():
    if "MicroDuck" in task.task_id:
        print(task.task_id)

Train a Specific Locomotion Task

uv run train Mjlab-VelStand-Flat-MicroDuck \
    --env.scene.num-envs 4096 \
    --agent.max_iterations 5000

Train with Backlash Simulation

uv run train Mjlab-VelStand-Flat-Backlash-MicroDuck \
    --env.scene.num-envs 4096

Deploy a Trained Policy

uv run play Mjlab-VelStand-Flat-MicroDuck \
    --wandb-run-path myuser/microduck/rl-run-123

Task Architecture and Extensibility

Understanding the source structure helps extend the locomotion task library:

1. Registration — register_mjlab_task() in __init__.py binds task_id to factory functions and config classes
2. Environment Config — Each make_*_env_cfg() returns an EnvCfg dataclass specifying robot model, terrain, observations, and commands
3. RL Config — *RlCfg classes inherit from MJLab's base and define reward weights, curricula, and algorithm settings
4. Runner — MicroduckOnPolicyRunner handles training loop execution with symmetry-configuration cleanup

To add a new locomotion task, create:

• A new *_env_cfg.py module with your make_*_env_cfg() factory
• A matching *RlCfg class
• One register_mjlab_task() call in __init__.py

The backlash variant generates automatically if desired.

Key Source Files

File	Purpose
src/mjlab_microduck/tasks/__init__.py	Central registry and backlash variant generation
src/mjlab_microduck/tasks/microduck_velocity_env_cfg.py	Standard velocity task environment
src/mjlab_microduck/tasks/microduck_velstand_env_cfg.py	VelStand combined task
src/mjlab_microduck/tasks/microduck_standup_env_cfg.py	Inverted recovery task
src/mjlab_microduck/tasks/microduck_ground_pick_env_cfg.py	Manipulation task
src/mjlab_microduck/tasks/microduck_roller_slope_env_cfg.py	Roller slope navigation
src/mjlab_microduck/tasks/backlash.py	Backlash injection helper
src/mjlab_microduck/tasks/mdp.py	Shared rewards, observations, and events

Summary

• 18 base locomotion tasks across velocity, recovery, manipulation, roller, and acrobatic categories
• Flat and rough terrain variants for most tasks to train generalization
• Automatic backlash variants (±1° gear-play) for sim-to-real robustness
• Modular registration system in src/mjlab_microduck/tasks/__init__.py enabling rapid task prototyping
• Recommended starting point: Mjlab-VelStand-Flat-MicroDuck for balanced locomotion, Mjlab-Velocity-Flat-Backlash-MicroDuck for hardware-ready policies

Frequently Asked Questions

What is the difference between Velocity and VelStand tasks?

Velocity tasks train pure command tracking with recovery as an emergent property. VelStand tasks explicitly reward pose maintenance and include curriculum elements for fall recovery, producing more robust policies for real-world deployment. VelStand uses MicroduckVelStandRlCfg with distinct reward shaping compared to the base MicroduckRlCfg.

How do I train a policy for rough terrain?

Append -Rough- to the base task_id or pass rough=True to the config factory function. For example: uv run train Mjlab-Velocity-Rough-MicroDuck. The rough terrain variant adds randomized height fields with configurable frequency and amplitude parameters in the environment configuration.

When should I use backlash variants?

Use backlash variants when preparing policies for physical hardware deployment. The ±1° joint backlash and encoder noise model in backlash.py forces the policy to tolerate imprecise state estimation and mechanical play. Training without backlash first, then fine-tuning with backlash, often produces the best results.

Can I create custom locomotion tasks?

Yes. The modular architecture requires: (1) a new *_env_cfg.py file with your make_*_env_cfg() factory returning an EnvCfg, (2) a *RlCfg class extending MJLab's base RL configuration, and (3) a register_mjlab_task() call in __init__.py. The existing microduck_spin_env_cfg.py provides a concise template for acrobatic tasks.

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