/RL-ROBOT

Reinforcement Learning framework for Robotics

Primary LanguagePythonOtherNOASSERTION

RL-ROBOT

Robot

This repository provides a Reinforcement Learning framework in Python from the Machine Perception and Intelligent Robotics research group (MAPIR).

Reference: Towards a common implementation of reinforcement learning for multiple robotics tasks. ScienceDirect(updated)       Arxiv preprint

Requirements

Tested on Ubuntu 14.04, 16.04, and 18.04 (64 bits)

  • Python >= 3.4
  • numpy >= 1.11
  • matplotlib >= 1.5
  • tkinter sudo apt install python-tk

Try RL-ROBOT:

Architecture

import exp
import rlrobot

exp.ENVIRONMENT_TYPE = "MODEL"   # "VREP" for V-REP simulation
exp.TASK_ID = "wander_1k"
exp.FILE_MODEL = exp.TASK_ID + "_model"

exp.ALGORITHM = "TOSL"
exp.ACTION_STRATEGY = "QBIASSR"
 
exp.N_REPETITIONS = 1
exp.N_EPISODES = 1
exp.N_STEPS = 60 * 60

exp.DISPLAY_STEP = 500

rlrobot.run()

Full set of parameters available in exp.py.

V-REP settings:

Tested: V-REP PRO EDU V3.3.2 / V3.5.0

Scenarios

  1. Use default values of remoteApiConnections.txt

    portIndex1_port 		= 19997
portIndex1_debug 		= false
portIndex1_syncSimTrigger 	= true

  2. Activate threaded rendering (recommended): system/usrset.txt -> threadedRenderingDuringSimulation = 1

Recommended simulation settings for V-REP scenes:

  • Simulation step time: 50 ms (default)
  • Real-Time Simulation: Enabled
  • Multiplication factor: 3.00 (required CPU >= i3 3110m)

Execute V-REP (./vrep.sh on linux). File -> Open Scene -> <RL-ROBOT path>/vrep_scenes