Robust Multi-Agent Tasks

Robust Multi-Agent Tasks#

Robust Multi-Agent Tasks#

TasksRobust type

Robust State

Robust Action

Robust Reward

MA-Ant-2x4, 2x4d, 4x2, 4x1

✅

✅

✅

MA-HalfCheetah-2x3, 6x1

✅

✅

✅

MA-Hopper-3x1

✅

✅

✅

MA-Walker2d-2x3

✅

✅

✅

MA-Swimmer-2x1

✅

✅

✅

MA-Humanoid-9|8

✅

✅

✅

MA-HumanoidStandup-v4

✅

✅

✅

MA-Pusher-3p

✅

✅

✅

MA-Reacher-2x1

✅

✅

✅

Many-MA-Swimmer-10x2, 5x4, 6x1, 1x2

✅

✅

✅

Many-MA-Ant-2x3, 3x1

✅

✅

✅

CoupledHalfCheetah-p1p

✅

✅

✅

A Simple Example

import numpy as np
from robust_gymnasium.envs.robust_ma_mujoco.mamujoco_v1 import get_parts_and_edges
from robust_gymnasium.envs.robust_ma_mujoco import mujoco_multi
from robust_gymnasium.configs.robust_setting import get_config

# Parse configuration
args = get_config().parse_args()

# Get nodes and edges for partitioning
unpartioned_nodes, edges, global_nodes = get_parts_and_edges('Ant', None)

# Define partitioned nodes and agent factorization
partioned_nodes = [
    (unpartioned_nodes[0][0],), (unpartioned_nodes[0][1],),
    (unpartioned_nodes[0][2],), (unpartioned_nodes[0][3],),
    (unpartioned_nodes[0][4],), (unpartioned_nodes[0][5],),
    (unpartioned_nodes[0][6],), (unpartioned_nodes[0][7],)
]
my_agent_factorization = {
    "partition": partioned_nodes,
    "edges": edges,
    "globals": global_nodes
}

# Initialize environment
env = mujoco_multi.MultiAgentMujocoEnv(
    'Ant', '8x1', agent_factorization=my_agent_factorization, render_mode="human"
)

# Supported environments:
# - MA-Ant-2x4, 2x4d, 4x2, 4x1
# - MA-HalfCheetah-2x3, 6x1
# - MA-Hopper-3x1
# - MA-Walker2d-2x3
# - MA-Swimmer-2x1
# - MA-Humanoid-9|8
# - MA-HumanoidStandup-v4
# - MA-Pusher-3p
# - MA-Reacher-2x1
# - Many-MA-Swimmer-10x2, 5x4, 6x1, 1x2
# - Many-MA-Ant-2x3, 3x1
# - CoupledHalfCheetah-p1p

# Reset environment and verify action mapping
observation, info = env.reset(seed=42)
action = env.single_agent_env.action_space.sample()
assert (
    action == env.map_local_actions_to_global_action(
        env.map_global_action_to_local_actions(action)
    )
).all()

# Generate actions for each agent
actions = []
n_actions = max([len(l) for l in my_agent_factorization['partition']])
for agent_id in range(8):
    avail_actions = [1.0]  # Example available actions
    avail_actions_ind = np.nonzero(avail_actions)[0]
    action = np.random.uniform(-1.0, 1.0, n_actions)
    actions.append(action)

# Predefined agent actions
re_actions = {
    'agent_0': [-0.34627497], 'agent_1': [0.13449878], 'agent_2': [-0.02019556],
    'agent_3': [0.5044606], 'agent_4': [0.8911963], 'agent_5': [-0.91603947],
    'agent_6': [-0.8271199], 'agent_7': [0.7132667]
}

# Run simulation
for i in range(1000):
    robust_input = {
        "action": re_actions,
        "robust_type": "state",
        "robust_config": args,
    }
    observation, reward, terminated, truncated, info = env.step(re_actions)
    env.render()
    if terminated or truncated:
        observation, info = env.reset()
env.close()