Implementation:Google deepmind Dm control Fruitfly V2

Knowledge Sources	Google_deepmind_Dm_control
Domains	Robotics Simulation, Neuroethology, Biomechanical Modeling
Last Updated	2026-02-15 04:00 GMT

Overview

The fruitfly_v2 module implements a detailed fruit fly (Drosophila) walker model with configurable body parts (legs, wings, mouth, antennae), adhesion actuators, binocular vision cameras, and rich sensory observables for neuroethology research.

Description

The FruitFly class extends legacy_base.Walker and loads a detailed MJCF model from an XML file. During its _build method, it performs several configuration steps: it optionally retracts unused body parts (legs, wings, mouth, antennae) by computing rest quaternions from joint springrefs and removing associated actuators, tendons, and joints; it configures wing stroke plane and body pitch angles for flight via quaternion transformations; it sets up actuator dynamics filters for both joint and adhesion actuators; and it builds action-class index maps (_ctrl_indices, _action_indices) that map environment actions to MuJoCo controls organized by body part category (adhesion, head, mouth, antennae, wings, abdomen, legs, user).

Several helper functions at module level provide quaternion math utilities (neg_quat, mul_quat, rot_vec_quat, mul_jac_t_vec) and body frame transformations (change_body_frame, body_quat_from_springrefs). The apply_action method maps action-class indices to MuJoCo control vectors, while get_action_spec constructs a bounded action specification from actuator control ranges.

The FruitFlyObservables class provides a rich set of sensory observables: thorax/abdomen height, world z-axis in multiple body frames (thorax, abdomen, head, hover up direction), force/touch/accelerometer/gyro/velocimeter sensors with buffered temporal aggregation, actuator activation, binocular eye cameras with configurable field of view and resolution, appendage positions in the egocentric frame, and self-contact force detection. Observable groups are organized into vestibular, proprioception, and orientation categories.

Usage

Use this module to create a fruit fly walker for locomotion, flight, or multi-modal sensory integration experiments. The walker is highly configurable: body parts can be independently enabled or disabled, timesteps can be adjusted, and additional user actions can be added. It integrates with the dm_control composer framework for use in locomotion tasks.

Code Reference

Source Location

• Repository: Google_deepmind_Dm_control
• File: dm_control/locomotion/walkers/fruitfly_v2.py
• Lines: 1-675

Signature

class FruitFly(legacy_base.Walker):
    def _build(self,
               name: str = 'walker',
               use_legs: bool = True,
               use_wings: bool = False,
               use_mouth: bool = False,
               use_antennae: bool = False,
               joint_filter: float = 0.01,
               adhesion_filter: float = 0.01,
               body_pitch_angle: float = 47.5,
               stroke_plane_angle: float = 0.,
               physics_timestep: float = 1e-4,
               control_timestep: float = 2e-3,
               num_user_actions: int = 0,
               eye_camera_fovy: float = 150.,
               eye_camera_size: int = 32):
        ...

    def initialize_episode(self, physics, random_state): ...
    def apply_action(self, physics, action, random_state): ...
    def get_action_spec(self, physics): ...

class FruitFlyObservables(legacy_base.WalkerObservables):
    def __init__(self, walker, buffer_size, eye_camera_size): ...

Import

from dm_control.locomotion.walkers.fruitfly_v2 import FruitFly

I/O Contract

Inputs

Name	Type	Required	Description
name	str	No	Name of the walker (default: 'walker')
use_legs	bool	No	Whether to use or retract the legs (default: True)
use_wings	bool	No	Whether to use or retract the wings (default: False)
use_mouth	bool	No	Whether to use or retract the mouth (default: False)
use_antennae	bool	No	Whether to use the antennae (default: False)
joint_filter	float	No	Timescale of filter for joint actuators; 0 disables (default: 0.01)
adhesion_filter	float	No	Timescale of filter for adhesion actuators; 0 disables (default: 0.01)
body_pitch_angle	float	No	Body pitch angle for initial flight pose in degrees (default: 47.5)
stroke_plane_angle	float	No	Wing stroke plane angle in degrees (default: 0.0)
physics_timestep	float	No	Simulation timestep in seconds (default: 1e-4)
control_timestep	float	No	Controller timestep in seconds (default: 2e-3)
num_user_actions	int	No	Number of additional custom actions with range [-1, 1] (default: 0)
eye_camera_fovy	float	No	Vertical field of view of eye cameras in degrees (default: 150.0)
eye_camera_size	int	No	Size in pixels of eye cameras (default: 32)

Outputs

Name	Type	Description
FruitFly instance	FruitFly	A configured fruit fly walker entity for use in dm_control composer tasks
action_spec	dm_env.specs.BoundedArray	Bounded action specification returned by get_action_spec(physics)

Usage Examples

Basic Usage

from dm_control.locomotion.walkers.fruitfly_v2 import FruitFly

# Create a walking fruit fly (legs only)
walker = FruitFly(name='fly', use_legs=True, use_wings=False)

# Create a flying fruit fly
flying_fly = FruitFly(
    name='fly',
    use_legs=False,
    use_wings=True,
    body_pitch_angle=47.5,
    stroke_plane_angle=0.,
)

# Access walker properties
root = walker.root_body
thorax = walker.thorax
end_effectors = walker.end_effectors

Related Pages

• Implementation:Google_deepmind_Dm_control_Legacy_Base_Walker