Multiagent Mujoco v1.1.0

.gitignore

@@ -1,3 +1,4 @@
 #.idea/
 .DS_Store
 *.pyc
+__pycache__/

README.md

@@ -1,6 +1,12 @@
 ```diff
-+ New Addition: Multi-Agent Mujoco now supports Scenarios with Coupled Robots (using Tendons). 
-+ Check out instructions below.
+-  PLEASE NOTE: To use the legacy version of Multiagent Mujoco please check out tag v1.0
+```
+
+```diff
++ New Version: Multi-Agent MuJoCo is now at Version 1.1.0. Changes and added features are as follows:
++ Fixed a bug in action mapping in the step function (thanks go to Paul Barde). This fixes several unphysical mappings found in previous versions.
++ Multi-Agent MuJoCo can now be installed as a PIP package.
++ I am in the process of establishing comprehensive benchmarks of all Multi-Agent MuJoCO scenarios across a variety of RL algorithms. 
 ```
 
 ```diff
@@ -17,20 +23,20 @@ Described in the paper [Deep Multi-Agent Reinforcement Learning for Decentralize
 
 # Installation
 
-**Note: You require OpenAI Gym Version 0.10.8 and Mujoco 2.0**
+**Note: You require OpenAI Gym Version 0.10.8 and Mujoco 2.1**
 
 Simply clone this repository and put ./src on your PYTHONPATH.
 To render, please also set the following environment variables:
 
 ```
-LD_LIBRARY_PATH=${HOME}/.mujoco/mujoco200/bin;
+LD_LIBRARY_PATH=${HOME}/.mujoco/mujoco210/bin;
 LD_PRELOAD=/usr/lib/x86_64-linux-gnu/libGLEW.so
 ```
 
 # Example
 
 ```python
-from src.multiagent_mujoco.mujoco_multi import MujocoMulti
+from multiagent_mujoco.mujoco_multi import MujocoMulti
 import numpy as np
 import time
 

src/multiagent_mujoco/mujoco_multi.py → multiagent_mujoco/mujoco_multi.py

@@ -30,10 +30,10 @@ class MujocoMulti(MultiAgentEnv):
 
     def __init__(self, batch_size=None, **kwargs):
         super().__init__(batch_size, **kwargs)
-        self.scenario = kwargs["env_args"]["scenario"] # e.g. Ant-v2
-        self.agent_conf = kwargs["env_args"]["agent_conf"] # e.g. '2x3'
+        self.scenario = kwargs["env_args"]["scenario"]  # e.g. Ant-v2
+        self.agent_conf = kwargs["env_args"]["agent_conf"]  # e.g. '2x3'
 
-        self.agent_partitions, self.mujoco_edges, self.mujoco_globals  = get_parts_and_edges(self.scenario,
+        self.agent_partitions, self.mujoco_edges, self.mujoco_globals = get_parts_and_edges(self.scenario,
                                                                                              self.agent_conf)
 
         self.n_agents = len(self.agent_partitions)
@@ -107,9 +107,18 @@ def __init__(self, batch_size=None, **kwargs):
 
     def step(self, actions):
 
-        # need to remove dummy actions that arise due to unequal action vector sizes across agents
-        flat_actions = np.concatenate([actions[i][:self.action_space[i].low.shape[0]] for i in range(self.n_agents)])
-        obs_n, reward_n, done_n, info_n = self.wrapped_env.step(flat_actions)
+        # we need to map actions back into MuJoCo action space
+        env_actions = np.zeros((sum([self.action_space[i].low.shape[0] for i in range(self.n_agents)]),)) + np.nan
+        for a, partition in enumerate(self.agent_partitions):
+            for i, body_part in enumerate(partition):
+                if env_actions[body_part.act_ids] == env_actions[body_part.act_ids]:
+                    raise Exception("FATAL: At least one env action is doubly defined!")
+                env_actions[body_part.act_ids] = actions[a][i]
+
+        if np.isnan(env_actions).any():
+            raise Exception("FATAL: At least one env action is undefined!")
+
+        obs_n, reward_n, done_n, info_n = self.wrapped_env.step(env_actions)
         self.steps += 1
 
         info = {}

setup.py

@@ -0,0 +1,50 @@
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+from setuptools import setup
+
+description = """MAMuJoCo - Multi-Agent MuJoCo
+
+Benchmark for Continuous Multi-Agent Robotic Control, based on OpenAI's Mujoco Gym environments.
+Described in the paper Deep Multi-Agent Reinforcement Learning for Decentralized Continuous Cooperative Control 
+by Christian Schroeder de Witt, Bei Peng, Pierre-Alexandre Kamienny, Philip Torr, Wendelin Böhmer and Shimon Whiteson, 
+Torr Vision Group and Whiteson Research Lab, University of Oxford, 2020
+
+Contact CSDW at [email protected]
+"""
+
+extras_deps = {
+      "dev": [
+            "pre-commit>=2.0.1",
+            "black>=19.10b0",
+            "flake8>=3.7",
+            "flake8-bugbear>=20.1",
+      ],
+}
+
+setup(
+      name="MAMujoco",
+      version="1.1.0",
+      description="MAMuJoCo - Multi-Agent MuJoCo.",
+      long_description=description,
+      author="Christian Schroeder de Witt",
+      author_email="[email protected]",
+      license="Apache 2.0 License",
+      keywords="Robotics, MuJoCo, Multi-Agent Reinforcement Learning",
+      url="https://github.com/schroederdewitt/multiagent_mujoco",
+      packages=[
+            "multiagent_mujoco",
+            "multiagent_mujoco.assets",
+      ],
+      extras_require=extras_deps,
+      install_requires=[
+            "numpy>=1.22.3",
+            "gym==0.10.8",
+            "mujoco-py>=2.1.2.14",
+            "scipy>=1.8.0",
+            "Jinja2>=3.0.3",
+            "glfw>=2.5.1",
+            "Cython>=0.29.28"
+      ],
+)