Implementation:Google deepmind Dm control Soccer Load

Metadata
Knowledge Sources	dm_control
Domains	Multi-Agent Reinforcement Learning, Environment Construction
Last Updated	2026-02-15 00:00 GMT

Overview

Concrete tool for constructing a complete N-versus-N multi-agent MuJoCo soccer environment with a single function call, handling player creation, pitch sizing, ball selection, and task wiring.

Description

The soccer.load function is the primary public entry point for the dm_control soccer environment. It performs the following steps:

1. Validates team_size (must be 1--11).
2. Sets default pitch size limits (min_size=(32, 24), max_size=(48, 36) for non-humanoid; computed from per-player area constants for humanoid walkers).
3. Selects ball type: SoccerBall() for boxhead/ant, regulation_soccer_ball() for humanoid.
4. Optionally adjusts goal size to MINI_FOOTBALL_GOAL_SIZE for humanoid walkers.
5. Selects Task or MultiturnTask based on terminate_on_goal.
6. Builds players via _make_players(team_size, walker_type).
7. Constructs a RandomizedPitch with the computed size bounds.
8. Wraps everything in a composer.Environment with the given time_limit and random_state.

Usage

This is the recommended way to create a soccer environment. All other components (walkers, pitch, task) are assembled automatically.

Code Reference

Attribute	Value
Source Location	dm_control/locomotion/soccer/__init__.py, lines 92--152
Signature	def load(team_size, time_limit=45.0, random_state=None, disable_walker_contacts=False, enable_field_box=False, keep_aspect_ratio=False, terminate_on_goal=True, walker_type=WalkerType.BOXHEAD)
Import	from dm_control.locomotion import soccer

I/O Contract

Inputs:

Parameter	Type	Description
team_size	int	Number of players per team (1--11).
time_limit	float	Maximum episode duration in seconds. Default 45.0.
random_state	int, np.random.RandomState, or None	Random seed or RNG instance. None uses a platform-dependent seed.
disable_walker_contacts	bool	Disable physical contacts between walkers. Default False.
enable_field_box	bool	Enable a physical bounding box for the ball. Default False.
keep_aspect_ratio	bool	Maintain constant pitch aspect ratio across randomisations. Default False.
terminate_on_goal	bool	If True, use Task (episode ends on goal). If False, use MultiturnTask. Default True.
walker_type	WalkerType	Walker morphology for all players. Default WalkerType.BOXHEAD.

Outputs:

Return	Type	Description
environment	composer.Environment	A fully configured multi-agent RL environment that implements the dm_env.Environment interface.

Usage Examples

from dm_control.locomotion import soccer

# 2v2 boxhead environment with default settings.
env = soccer.load(team_size=2)

# 3v3 humanoid environment, continuous play, 90 second episodes.
env = soccer.load(
    team_size=3,
    time_limit=90.0,
    walker_type=soccer.WalkerType.HUMANOID,
    terminate_on_goal=False,
)

# 1v1 with deterministic seed and no walker contacts.
env = soccer.load(
    team_size=1,
    random_state=42,
    disable_walker_contacts=True,
)

# Inspect the environment.
timestep = env.reset()
print(type(timestep))  # <class 'dm_env._environment.TimeStep'>
print(len(timestep.observation))  # 2 * team_size observation dicts

action_specs = env.action_spec()
print(len(action_specs))  # 2 * team_size action specs

Related Pages

• Principle:Google_deepmind_Dm_control_Soccer_Environment_Loading
• Environment:Google_deepmind_Dm_control_Python_MuJoCo_Runtime