Multiagent Follow Envirnoment

Object Oriented MultiAgent Environment Using Pybullet

There is a lag before it is loaded.

🌱 Installation

pip install -e .

🌱 User Guide

🐋 1. objects in the environment

Every objects in the enviroment is in the environment.objects whose type is dictionary.

# for example
self.objects = {'target':[PhysicalObjects(), PhysicalObjects()],
                'agent':[Agent(), Agent()], 
                'obstacle':[PhysicalObjects(), PhysicalObjects(), PhysicalObjects()]}

You can easily calculate an interaction between two objects in the environment

Every objects in the environment inherits PhysicalObjects Class which is the base class for the moving objects.

member variables of the object	description	type
self.position	current position of the object	list[3]
self.velocity	current velocity of the object	list[3]
self.max_speed	maximum speed, the velocity will be bounded	float
self.acc	acceleration	float
self.speed_decreasing	effect like friction	float
self.pid	pybullet id. you can get it by calling p.loadURDF( )	int
self.safe_boundary	to calculate collision between two agents	float
self.move_kind	the movement strategy for the current timestep	str

Of course you can add more variables if you want!

🐋 2. Design strategy of the agent

Agent is also the PhysicalObject and there are additional functions for Agent.

With these functions you can implement reinforcement learning code easily

# Example :  FollowAvoid Scenario - Agent Class
# Take action 
def take_action(self, action):
    if action < 5 :
        p.applyExternalForce(self.pid, -1, 
                                forceObj = [Agent.dx[action]*self.acc, Agent.dy[action]*self.acc, 0],
                                posObj=self.position,
                                flags=p.WORLD_FRAME) 
    elif action == 5 :
        p.resetBaseVelocity(self.pid, [0,0,0])
    else:
        raise ValueError("Undefined action %d" %action)

# compute relative position with other PhysicalObject object
def relative_position(self, other):
    reltaive = [other.position[i] - self.position[i] for i in range(3)]
    return reltaive
    
...

🐋 3. Reinforcement Learning

By giving the position of the obstacles as the observaiton of the agent, you can implement Avoid Obstacles RL task.

• Green : Agent
• White : Obstacle

🌱 Questions

If there is some issues or questions please Submit new Issue.