[Task] VMAS football (#166)

* football

* football

* fix

* amend

benchmarl/conf/task/vmas/football.yaml

@@ -0,0 +1,66 @@
+defaults:
+  - vmas_football_config
+  - _self_
+
+
+max_steps: 500
+
+# Agents config
+n_blue_agents: 3
+n_red_agents: 3
+ai_red_agents: True  # What agents should be learning and what controlled by the heuristic (ai)
+
+# When you have 5 blue agents, physical differences can be introduced:
+# 1 goalkeeper -> slow and big
+# 2 defenders -> normal size and speed (agent_size, u_multiplier, max_speed)
+# 2 attackers -> small and fast
+physically_different: False
+
+# Agent spawning
+spawn_in_formation: False
+only_blue_formation: True  # Only spawn blue agents in formation
+formation_agents_per_column: 2
+randomise_formation_indices: False  # If False, each agent will always be in the same formation spot
+formation_noise: 0.2  # Noise on formation positions
+
+# Ai config
+n_traj_points: 0  # Number of spline trajectory points to plot for heuristic (ai) agents
+ai_strength: 1.0  # The speed of the ai 0<=x<=1
+ai_decision_strength: 1.0  # The decision strength of the ai 0<=x<=1
+ai_precision_strength: 1.0  # The precision strength of the ai 0<=x<=1
+
+# Task sizes
+agent_size: 0.025
+goal_size: 0.35
+goal_depth: 0.1
+pitch_length: 3.0
+pitch_width: 1.5
+ball_mass: 0.25
+ball_size: 0.02
+
+# Actions
+u_multiplier: 0.1
+
+# Actions shooting
+enable_shooting: False  # Whether to enable an extra 2 actions (for rotation and shooting). Only available for non-ai agents
+u_rot_multiplier: 0.0003
+u_shoot_multiplier: 0.6
+shooting_radius: 0.08
+shooting_angle: 1.5708
+
+# Speeds
+max_speed: 0.15
+ball_max_speed: 0.3
+
+# Rewards
+dense_reward: True
+pos_shaping_factor_ball_goal: 10.0  # Reward for moving the ball towards the opponents' goal. This can be annealed in a curriculum.
+pos_shaping_factor_agent_ball: 0.1  # Reward for moving the closest agent to the ball in a team closer to it.
+# This reward does not trigger if the agent is less than distance_to_ball_trigger from the ball or the ball is moving
+distance_to_ball_trigger: 0.4
+scoring_reward: 100.0  # Discrete reward for scoring
+
+# Observations
+observe_teammates: True
+observe_adversaries: True
+dict_obs: False

benchmarl/environments/vmas/common.py

@@ -36,6 +36,7 @@ class VmasTask(Task):
     BUZZ_WIRE = None
     FLOCKING = None
     DISCOVERY = None
+    FOOTBALL = None
     SIMPLE_ADVERSARY = None
     SIMPLE_CRYPTO = None
     SIMPLE_PUSH = None

benchmarl/environments/vmas/football.py

@@ -0,0 +1,66 @@
+#  Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+#  This source code is licensed under the license found in the
+#  LICENSE file in the root directory of this source tree.
+#
+
+from dataclasses import dataclass, MISSING
+
+
+@dataclass
+class TaskConfig:
+    max_steps: int = MISSING
+
+    # Agents config
+    n_blue_agents: int = MISSING
+    n_red_agents: int = MISSING
+    ai_red_agents: bool = MISSING
+    physically_different: bool = MISSING
+
+    # Agent spawning
+    spawn_in_formation: bool = MISSING
+    formation_agents_per_column: int = MISSING
+    randomise_formation_indices: bool = MISSING
+    only_blue_formation: bool = MISSING
+    formation_noise: float = MISSING
+
+    # Opponent heuristic config
+    n_traj_points: int = MISSING
+    ai_strength: float = MISSING
+    ai_decision_strength: float = MISSING
+    ai_precision_strength: float = MISSING
+
+    # Task sizes
+    agent_size: float = MISSING
+    goal_size: float = MISSING
+    goal_depth: float = MISSING
+    pitch_length: float = MISSING
+    pitch_width: float = MISSING
+    ball_mass: float = MISSING
+    ball_size: float = MISSING
+
+    # Actions
+    u_multiplier: float = MISSING
+
+    # Actions shooting
+    enable_shooting: bool = MISSING
+    u_rot_multiplier: float = MISSING
+    u_shoot_multiplier: float = MISSING
+    shooting_radius: float = MISSING
+    shooting_angle: float = MISSING
+
+    # Speeds
+    max_speed: float = MISSING
+    ball_max_speed: float = MISSING
+
+    # Rewards
+    dense_reward: bool = MISSING
+    pos_shaping_factor_ball_goal: float = MISSING
+    pos_shaping_factor_agent_ball: float = MISSING
+    distance_to_ball_trigger: float = MISSING
+    scoring_reward: float = MISSING
+
+    # Observations
+    observe_teammates: bool = MISSING
+    observe_adversaries: bool = MISSING
+    dict_obs: bool = MISSING