francesco-mannella / ArmSim

ArmSim

A simple gym environment using pybox2d as the physics engine and matplotlib for graphics.

Table of contents

• Install
• Basic usage

Install

1. Download de ArmSim repo:

   git clone https://github.com/francesco-mannella/ArmSim.git
   

2. Install the ArmSim package: cd ArmSim pip install -e .

Basic usage

One-arm scenario

import gym
import ArmSim

env = gym.make('ArmSimOneArm-v0')

for t in range(10):  
  env.render()
  observation = env.step(env.action_space.sample())

rendering

The two possible values of the argument to be passed to env.render() are:

• "human": open a matplotlib figure and update it at each call.
• "offline": save a frame into a png file at each call. Files are saved into the local folder 'frames'. This folder is created if it does not exist.

Example in test.py

Actions

The action attribute of env.step must be a vector of 7 joint positions in radiants. The first 7 joints have a range between -Pi/2 and +Pi/2. The two gripper joints have a range between 0 and +Pi/2. They are also coupled so that the second joint will be at most twice the angle of the first one.

index joint name 0 Ground_to_Arm1 1 Arm1_to_Arm2 2 Arm2_to_Arm3 3 Arm3_to_Claw20 4 Claw20_to_Claw21

Observations

The observation object returned by env.step is a dictionary:

• observation["JOINT_POSITIONS"] is a vector containing the current angles of the 7:w joints
• observation["TOUCH_SENSORS"] is a vector containing the current touch intensity at the four touch sensors (see figure below)
• observation["OBJ_POSITION"] coordinates of the center of mass of the external object

Reward

The reward value returned by env.step is always put to 0.

Done

The done value returned by env.step is always set to False.

Info

The info value returned by env.step is always set to an empy set {}. .