Safe Multi-Agent Robosuite Benchmark

Safe Multi-Agent Robosuite is an extension of Robosuite . We split the control over a robot across multiple controllers of its joints---one or more per controller. For example, the Lift task comes with 3 variants: 2 four-dimensional agents (2x4 Lift), 4 two-dimensional agents (4x2 Lift), and 8 one-dimensional agents (8x1 Lift). (This repository is under actively development. We appreciate any constructive comments and suggestions)

Safe MARobosuite tasks are fully cooperative, partialy observable, continuous, and safety-aware. Its multi-agency makes it a compatible framework for training modular robots which are built of multiple, robust parts, refer to the work. We adopt the reward setting from Robosuite.

Figure.1 Example tasks in Safe MARobosuite Environment. (a): Safe 2x4-Lift, (b): Safe 4x2-Lift, (c): Safe 8x1-Lift, (d): Safe 2x4-Stack, (e): Safe 4x2-Stack, (f): Safe 8x1-Stack, (g): Safe 14x1-TwoArmPegInHole, (h): Safe 2x7-TwoArmPegInHole. Body parts of different colours of robots are controlled by different agents. Agents jointly learn to manipulate the robot, while avoiding crashing into unsafe red areas.

Installation

• Install Robosuite accoring to Robosuite and MuJoCo website.
• clone safety multi-agent mujoco to the env path.  

LD_LIBRARY_PATH=${HOME}/.mujoco/mujoco200/bin;
LD_PRELOAD=/usr/lib/x86_64-linux-gnu/libGLEW.so

pip install -e .

Quick Start

import numpy as np
import robosuite as suite

# create environment instance
env = suite.make(
    env_name="Lift", 
    robots="Panda",  
    has_renderer=True,
    has_offscreen_renderer=False,
    use_camera_obs=False,
)

# reset the environment
env.reset()

for i in range(1000):
    action = np.random.randn(env.robots[0].dof) # sample random action
    obs, reward, cost, done, info = env.step(action)  # take action in the environment
    env.render()  # render on display

Tasks

    Lift
    env_args = {"scenario": "Lift",
                  "smarobosuite_robots": "panda"
                  "agent_conf": "2x4",
                  "agent_obsk": 1,
                  "episode_limit": 1000}
                  
    env_args = {"scenario": "Lift",
                  "smarobosuite_robots": "panda"
                  "agent_conf": "4x2",
                  "agent_obsk": 1,
                  "episode_limit": 1000}
                  
    env_args = {"scenario": "Lift",
                  "smarobosuite_robots": "panda"
                  "agent_conf": "8x1",
                  "agent_obsk": 1,
                  "episode_limit": 1000}
                  
    Stack
    env_args = {"scenario": "Stack",
                  "smarobosuite_robots": "panda"
                  "agent_conf": "2x4",
                  "agent_obsk": 1,
                  "episode_limit": 1000}
                  
    env_args = {"scenario": "Stack",
                  "smarobosuite_robots": "panda"
                  "agent_conf": "4x2",
                  "agent_obsk": 1,
                  "episode_limit": 1000}
                  
    env_args = {"scenario": "Stack",
                  "smarobosuite_robots": "panda"
                  "agent_conf": "8x1",
                  "agent_obsk": 1,
                  "episode_limit": 1000}
                  
    TwoArmPegInHole
                  
    env_args = {"scenario": "TwoArmPegInHole",
                  "smarobosuite_robots": ["panda", "panda"],
                  "agent_conf": "2x7",
                  "agent_obsk": 1,
                  "episode_limit": 1000}