In this new ROS Project you are going to learn Step-by-Step how to create a robot cube that moves and that it learns to move using OpenAI environment.
In this fifth and last video of the cube series, we create the Robot environment for OpenAI Gym for the moving cube. You will learn all the details on why we do it like this and we finally execute the script to make the robot cube learn how to walk in one direction.
For any suggestion on the next AI project that we could do, please leave us a comment, we will be happy to hear your ideas :).
We’ll start by creating a file called old_way_moving_cube_env.py under the my_moving_cube_training_pkg/script directory with the following content
import gym
import rospy
import time
import numpy as np
import math
import copy
from gym import utils, spaces
import numpy
from std_msgs.msg import Float64
from sensor_msgs.msg import JointState
from rosgraph_msgs.msg import Clock
from nav_msgs.msg import Odometry
from gazebo_connection import GazeboConnection
from controllers_connection import ControllersConnection
from gym.utils import seeding
from gym.envs.registration import register
from geometry_msgs.msg import Point
from tf.transformations import euler_from_quaternion
reg = register(
id='MyOldMovingCube-v0',
entry_point='old_way_moving_cube_env:MyOldMovingCubeEnv',
timestep_limit=1000,
)
class MyOldMovingCubeEnv(gym.Env):
def __init__(self):
number_actions = rospy.get_param('/moving_cube/n_actions')
self.action_space = spaces.Discrete(number_actions)
self._seed()
#get configuration parameters
self.init_roll_vel = rospy.get_param('/moving_cube/init_roll_vel')
# Actions
self.roll_speed_fixed_value = rospy.get_param('/moving_cube/roll_speed_fixed_value')
self.roll_speed_increment_value = rospy.get_param('/moving_cube/roll_speed_increment_value')
self.start_point = Point()
self.start_point.x = rospy.get_param("/moving_cube/init_cube_pose/x")
self.start_point.y = rospy.get_param("/moving_cube/init_cube_pose/y")
self.start_point.z = rospy.get_param("/moving_cube/init_cube_pose/z")
# Done
self.max_pitch_angle = rospy.get_param('/moving_cube/max_pitch_angle')
# Rewards
self.move_distance_reward_weight = rospy.get_param("/moving_cube/move_distance_reward_weight")
self.y_linear_speed_reward_weight = rospy.get_param("/moving_cube/y_linear_speed_reward_weight")
self.y_axis_angle_reward_weight = rospy.get_param("/moving_cube/y_axis_angle_reward_weight")
self.end_episode_points = rospy.get_param("/moving_cube/end_episode_points")
# stablishes connection with simulator
self.gazebo = GazeboConnection()
self.controllers_list = ['joint_state_controller',
'inertia_wheel_roll_joint_velocity_controller'
]
self.controllers_object = ControllersConnection(namespace="moving_cube",
controllers_list=self.controllers_list)
self.gazebo.unpauseSim()
self.controllers_object.reset_controllers()
self.check_all_sensors_ready()
rospy.Subscriber("/moving_cube/joint_states", JointState, self.joints_callback)
rospy.Subscriber("/moving_cube/odom", Odometry, self.odom_callback)
self._roll_vel_pub = rospy.Publisher('/moving_cube/inertia_wheel_roll_joint_velocity_controller/command', Float64, queue_size=1)
self.check_publishers_connection()
self.gazebo.pauseSim()
def _seed(self, seed=None): #overriden function
self.np_random, seed = seeding.np_random(seed)
return [seed]
def _step(self, action):#overriden function
self.gazebo.unpauseSim()
self.set_action(action)
self.gazebo.pauseSim()
obs = self._get_obs()
done = self._is_done(obs)
info = {}
reward = self.compute_reward(obs, done)
simplified_obs = self.convert_obs_to_state(obs)
return simplified_obs, reward, done, info
def _reset(self):
self.gazebo.unpauseSim()
self.controllers_object.reset_controllers()
self.check_all_sensors_ready()
self.set_init_pose()
self.gazebo.pauseSim()
self.gazebo.resetSim()
self.gazebo.unpauseSim()
self.controllers_object.reset_controllers()
self.check_all_sensors_ready()
self.gazebo.pauseSim()
self.init_env_variables()
obs = self._get_obs()
simplified_obs = self.convert_obs_to_state(obs)
return simplified_obs
def init_env_variables(self):
"""
Inits variables needed to be initialised each time we reset at the start
of an episode.
:return:
"""
self.total_distance_moved = 0.0
self.current_y_distance = self.get_y_dir_distance_from_start_point(self.start_point)
self.roll_turn_speed = rospy.get_param('/moving_cube/init_roll_vel')
def _is_done(self, observations):
pitch_angle = observations[3]
if abs(pitch_angle) > self.max_pitch_angle:
rospy.logerr("WRONG Cube Pitch Orientation==>" + str(pitch_angle))
done = True
else:
rospy.logdebug("Cube Pitch Orientation Ok==>" + str(pitch_angle))
done = False
return done
def set_action(self, action):
# We convert the actions to speed movements to send to the parent class CubeSingleDiskEnv
if action == 0:# Move Speed Wheel Forwards
self.roll_turn_speed = self.roll_speed_fixed_value
elif action == 1:# Move Speed Wheel Backwards
self.roll_turn_speed = self.roll_speed_fixed_value
elif action == 2:# Stop Speed Wheel
self.roll_turn_speed = 0.0
elif action == 3:# Increment Speed
self.roll_turn_speed += self.roll_speed_increment_value
elif action == 4:# Decrement Speed
self.roll_turn_speed -= self.roll_speed_increment_value
# We clamp Values to maximum
rospy.logdebug("roll_turn_speed before clamp=="+str(self.roll_turn_speed))
self.roll_turn_speed = numpy.clip(self.roll_turn_speed,
-1*self.roll_speed_fixed_value,
self.roll_speed_fixed_value)
rospy.logdebug("roll_turn_speed after clamp==" + str(self.roll_turn_speed))
# We tell the OneDiskCube to spin the RollDisk at the selected speed
self.move_joints(self.roll_turn_speed)
def _get_obs(self):
"""
Here we define what sensor data defines our robots observations
To know which Variables we have acces to, we need to read the
MyCubeSingleDiskEnv API DOCS
:return:
"""
# We get the orientation of the cube in RPY
roll, pitch, yaw = self.get_orientation_euler()
# We get the distance from the origin
y_distance = self.get_y_dir_distance_from_start_point(self.start_point)
# We get the current speed of the Roll Disk
current_disk_roll_vel = self.get_roll_velocity()
# We get the linear speed in the y axis
y_linear_speed = self.get_y_linear_speed()
cube_observations = [
round(current_disk_roll_vel, 0),
round(y_distance, 1),
round(roll, 1),
round(pitch, 1),
round(y_linear_speed,1),
round(yaw, 1),
]
return cube_observations
def get_orientation_euler(self):
# We convert from quaternions to euler
orientation_list = [self.odom.pose.pose.orientation.x,
self.odom.pose.pose.orientation.y,
self.odom.pose.pose.orientation.z,
self.odom.pose.pose.orientation.w]
roll, pitch, yaw = euler_from_quaternion(orientation_list)
return roll, pitch, yaw
def get_roll_velocity(self):
# We get the current joint roll velocity
roll_vel = self.joints.velocity[0]
return roll_vel
def get_y_linear_speed(self):
# We get the current joint roll velocity
y_linear_speed = self.odom.twist.twist.linear.y
return y_linear_speed
def get_y_dir_distance_from_start_point(self, start_point):
"""
Calculates the distance from the given point and the current position
given by odometry. In this case the increase or decrease in y.
:param start_point:
:return:
"""
y_dist_dir = self.odom.pose.pose.position.y - start_point.y
return y_dist_dir
def compute_reward(self, observations, done):
if not done:
y_distance_now = observations[1]
delta_distance = y_distance_now - self.current_y_distance
rospy.logdebug("y_distance_now=" + str(y_distance_now)+", current_y_distance=" + str(self.current_y_distance))
rospy.logdebug("delta_distance=" + str(delta_distance))
reward_distance = delta_distance * self.move_distance_reward_weight
self.current_y_distance = y_distance_now
y_linear_speed = observations[4]
rospy.logdebug("y_linear_speed=" + str(y_linear_speed))
reward_y_axis_speed = y_linear_speed * self.y_linear_speed_reward_weight
# Negative Reward for yaw different from zero.
yaw_angle = observations[5]
rospy.logdebug("yaw_angle=" + str(yaw_angle))
# Worst yaw is 90 and 270 degrees, best 0 and 180. We use sin function for giving reward.
sin_yaw_angle = math.sin(yaw_angle)
rospy.logdebug("sin_yaw_angle=" + str(sin_yaw_angle))
reward_y_axis_angle = -1 * abs(sin_yaw_angle) * self.y_axis_angle_reward_weight
# We are not intereseted in decimals of the reward, doesnt give any advatage.
reward = round(reward_distance, 0) + round(reward_y_axis_speed, 0) + round(reward_y_axis_angle, 0)
rospy.logdebug("reward_distance=" + str(reward_distance))
rospy.logdebug("reward_y_axis_speed=" + str(reward_y_axis_speed))
rospy.logdebug("reward_y_axis_angle=" + str(reward_y_axis_angle))
rospy.logdebug("reward=" + str(reward))
else:
reward = -1*self.end_episode_points
return reward
def joints_callback(self, data):
self.joints = data
def odom_callback(self, data):
self.odom = data
def check_all_sensors_ready(self):
self.check_joint_states_ready()
self.check_odom_ready()
rospy.logdebug("ALL SENSORS READY")
def check_joint_states_ready(self):
self.joints = None
while self.joints is None and not rospy.is_shutdown():
try:
self.joints = rospy.wait_for_message("/moving_cube/joint_states", JointState, timeout=1.0)
rospy.logdebug("Current moving_cube/joint_states READY=>" + str(self.joints))
except:
rospy.logerr("Current moving_cube/joint_states not ready yet, retrying for getting joint_states")
return self.joints
def check_odom_ready(self):
self.odom = None
while self.odom is None and not rospy.is_shutdown():
try:
self.odom = rospy.wait_for_message("/moving_cube/odom", Odometry, timeout=1.0)
rospy.logdebug("Current /moving_cube/odom READY=>" + str(self.odom))
except:
rospy.logerr("Current /moving_cube/odom not ready yet, retrying for getting odom")
return self.odom
def check_publishers_connection(self):
"""
Checks that all the publishers are working
:return:
"""
rate = rospy.Rate(10) # 10hz
while (self._roll_vel_pub.get_num_connections() == 0 and not rospy.is_shutdown()):
rospy.logdebug("No susbribers to _roll_vel_pub yet so we wait and try again")
try:
rate.sleep()
except rospy.ROSInterruptException:
# This is to avoid error when world is rested, time when backwards.
pass
rospy.logdebug("_base_pub Publisher Connected")
rospy.logdebug("All Publishers READY")
def move_joints(self, roll_speed):
joint_speed_value = Float64()
joint_speed_value.data = roll_speed
rospy.logdebug("Single Disk Roll Velocity>>" + str(joint_speed_value))
self._roll_vel_pub.publish(joint_speed_value)
self.wait_until_roll_is_in_vel(joint_speed_value.data)
def wait_until_roll_is_in_vel(self, velocity):
rate = rospy.Rate(10)
start_wait_time = rospy.get_rostime().to_sec()
end_wait_time = 0.0
epsilon = 0.1
v_plus = velocity + epsilon
v_minus = velocity - epsilon
while not rospy.is_shutdown():
joint_data = self.check_joint_states_ready()
roll_vel = joint_data.velocity[0]
rospy.logdebug("VEL=" + str(roll_vel) + ", ?RANGE=[" + str(v_minus) + ","+str(v_plus)+"]")
are_close = (roll_vel <= v_plus) and (roll_vel > v_minus)
if are_close:
rospy.logdebug("Reached Velocity!")
end_wait_time = rospy.get_rostime().to_sec()
break
rospy.logdebug("Not there yet, keep waiting...")
rate.sleep()
delta_time = end_wait_time- start_wait_time
rospy.logdebug("[Wait Time=" + str(delta_time)+"]")
return delta_time
def set_init_pose(self):
"""Sets the Robot in its init pose
"""
self.move_joints(self.init_roll_vel)
return True
def convert_obs_to_state(self,observations):
"""
Converts the observations used for reward and so on to the essentials for the robot state
In this case we only need the orientation of the cube and the speed of the disc.
The distance doesnt condition at all the actions
"""
disk_roll_vel = observations[0]
y_linear_speed = observations[4]
yaw_angle = observations[5]
state_converted = [disk_roll_vel, y_linear_speed, yaw_angle]
return state_converted
From line 22 to line 26 is the most important part of the file. You have to register your environment to OpenAI gym.
In the _get_obs() function in line 169 to line 198, the observation is measured from the simulation and will be used to generate a state.
The set_action() function in line 144 to line 166 executes the action decided by the q learning algorithm.
From line 88 to 188, the code defines what should the environment do when the algorithm runs into step or reset state.
The reward and checking if the training session is done or not is done in the __step() function.
In _reset() function, the environment will at first pause the gazebo simulation and reset the controller. In order to connect to gazebo and controller, we create a file called gazebo_connection.py under the same folder with the following content
#!/usr/bin/env python
import rospy
from std_srvs.srv import Empty
from gazebo_msgs.msg import ODEPhysics
from gazebo_msgs.srv import SetPhysicsProperties, SetPhysicsPropertiesRequest
from std_msgs.msg import Float64
from geometry_msgs.msg import Vector3
class GazeboConnection():
def __init__(self):
self.unpause = rospy.ServiceProxy('/gazebo/unpause_physics', Empty)
self.pause = rospy.ServiceProxy('/gazebo/pause_physics', Empty)
self.reset_proxy = rospy.ServiceProxy('/gazebo/reset_simulation', Empty)
# Setup the Gravity Controle system
service_name = '/gazebo/set_physics_properties'
rospy.logdebug("Waiting for service " + str(service_name))
rospy.wait_for_service(service_name)
rospy.logdebug("Service Found " + str(service_name))
self.set_physics = rospy.ServiceProxy(service_name, SetPhysicsProperties)
self.init_values()
# We always pause the simulation, important for legged robots learning
self.pauseSim()
def pauseSim(self):
rospy.wait_for_service('/gazebo/pause_physics')
try:
self.pause()
except rospy.ServiceException as e:
print ("/gazebo/pause_physics service call failed")
def unpauseSim(self):
rospy.wait_for_service('/gazebo/unpause_physics')
try:
self.unpause()
except rospy.ServiceException as e:
print ("/gazebo/unpause_physics service call failed")
def resetSim(self):
rospy.wait_for_service('/gazebo/reset_simulation')
try:
self.reset_proxy()
except rospy.ServiceException as e:
print ("/gazebo/reset_simulation service call failed")
def resetWorld(self):
rospy.wait_for_service('/gazebo/reset_world')
try:
self.reset_proxy()
except rospy.ServiceException as e:
print ("/gazebo/reset_world service call failed")
def init_values(self):
rospy.wait_for_service('/gazebo/reset_simulation')
try:
# reset_proxy.call()
self.reset_proxy()
except rospy.ServiceException as e:
print ("/gazebo/reset_simulation service call failed")
self._time_step = Float64(0.001)
self._max_update_rate = Float64(1000.0)
self._gravity = Vector3()
self._gravity.x = 0.0
self._gravity.y = 0.0
self._gravity.z = -9.81
self._ode_config = ODEPhysics()
self._ode_config.auto_disable_bodies = False
self._ode_config.sor_pgs_precon_iters = 0
self._ode_config.sor_pgs_iters = 50
self._ode_config.sor_pgs_w = 1.3
self._ode_config.sor_pgs_rms_error_tol = 0.0
self._ode_config.contact_surface_layer = 0.001
self._ode_config.contact_max_correcting_vel = 0.0
self._ode_config.cfm = 0.0
self._ode_config.erp = 0.2
self._ode_config.max_contacts = 20
self.update_gravity_call()
def update_gravity_call(self):
self.pauseSim()
set_physics_request = SetPhysicsPropertiesRequest()
set_physics_request.time_step = self._time_step.data
set_physics_request.max_update_rate = self._max_update_rate.data
set_physics_request.gravity = self._gravity
set_physics_request.ode_config = self._ode_config
rospy.logdebug(str(set_physics_request.gravity))
result = self.set_physics(set_physics_request)
rospy.logdebug("Gravity Update Result==" + str(result.success) + ",message==" + str(result.status_message))
self.unpauseSim()
def change_gravity(self, x, y, z):
self._gravity.x = x
self._gravity.y = y
self._gravity.z = z
self.update_gravity_call()
Then we create another file called controllers_connection.py with the following content
#!/usr/bin/env python
import rospy
import time
from controller_manager_msgs.srv import SwitchController, SwitchControllerRequest, SwitchControllerResponse
class ControllersConnection():
def __init__(self, namespace, controllers_list):
self.controllers_list = controllers_list
self.switch_service_name = '/'+namespace+'/controller_manager/switch_controller'
self.switch_service = rospy.ServiceProxy(self.switch_service_name, SwitchController)
def switch_controllers(self, controllers_on, controllers_off, strictness=1):
"""
Give the controllers you want to switch on or off.
:param controllers_on: ["name_controler_1", "name_controller2",...,"name_controller_n"]
:param controllers_off: ["name_controler_1", "name_controller2",...,"name_controller_n"]
:return:
"""
rospy.wait_for_service(self.switch_service_name)
try:
switch_request_object = SwitchControllerRequest()
switch_request_object.start_controllers = controllers_on
switch_request_object.start_controllers = controllers_off
switch_request_object.strictness = strictness
switch_result = self.switch_service(switch_request_object)
"""
[controller_manager_msgs/SwitchController]
int32 BEST_EFFORT=1
int32 STRICT=2
string[] start_controllers
string[] stop_controllers
int32 strictness
---
bool ok
"""
rospy.logdebug("Switch Result==>"+str(switch_result.ok))
return switch_result.ok
except rospy.ServiceException as e:
print (self.switch_service_name+" service call failed")
return None
def reset_controllers(self):
"""
We turn on and off the given controllers
:param controllers_reset: ["name_controler_1", "name_controller2",...,"name_controller_n"]
:return:
"""
reset_result = False
result_off_ok = self.switch_controllers(controllers_on = [],
controllers_off = self.controllers_list)
rospy.logdebug("Deactivated Controlers")
if result_off_ok:
rospy.logdebug("Activating Controlers")
result_on_ok = self.switch_controllers(controllers_on=self.controllers_list,
controllers_off=[])
if result_on_ok:
rospy.logdebug("Controllers Reseted==>"+str(self.controllers_list))
reset_result = True
else:
rospy.logdebug("result_on_ok==>" + str(result_on_ok))
else:
rospy.logdebug("result_off_ok==>" + str(result_off_ok))
return reset_result
def update_controllers_list(self, new_controllers_list):
self.controllers_list = new_controllers_list
Step 2. Start training
Now you have all the script you need for training, let’s create a launch file to launch the training under my_moving_cube_training_pkg/launch with the name start_training.launch
<launch>
<rosparam command="load" file="$(find my_moving_cube_training_pkg)/config/one_disk_walk_openai_params.yaml" />
<!-- Launch the training system -->
<node pkg="my_moving_cube_training_pkg" name="movingcube_gym" type="oldway_start_training.py" output="screen"/>
</launch>
If you already closed the simulation, please start it again from Simulations->select launch file->main.launch
Then run the following command to launch the training.
cd ~/catkin_ws
source devel/setup.bash
roslaunch my_moving_cube_training_pkg start_training.launch
You should see the cube robot now moving around to find the best way to move, you can play with different parameters to optimize the training.
If you are interested in this topic, please do not forget to check our OpenAI course at Robot Ignite Academy where you can learn how to create the gym environment for different robots!
In this new ROS Project you are going to learn Step-by-Step how to create a robot cube that moves and that it learns to move using OpenAI environment.
In this fourth video we continue we talk about the first script we need to do reinforcement learning with OpenAI. Specifically the main script where we import the Robot Environment that we will define in the next video and we will use Qlearn.
#!/usr/bin/env python
'''
Training code made by Ricardo Tellez <rtellez@theconstructsim.com>
Based on many other examples around Internet
Visit our website at www.theconstruct.ai
'''
import gym
import numpy
import time
import qlearn
from gym import wrappers
# ROS packages required
import rospy
import rospkg
# import our training environment
import old_way_moving_cube_env
if __name__ == '__main__':
rospy.init_node('movingcube_gym', anonymous=True, log_level=rospy.WARN)
# Create the Gym environment
env = gym.make('OldMovingCube-v0')
rospy.loginfo ( "Gym environment done")
# Set the logging system
rospack = rospkg.RosPack()
pkg_path = rospack.get_path('moving_cube_training_pkg')
outdir = pkg_path + '/training_results'
env = wrappers.Monitor(env, outdir, force=True)
rospy.loginfo ( "Monitor Wrapper started")
last_time_steps = numpy.ndarray(0)
# Loads parameters from the ROS param server
# Parameters are stored in a yaml file inside the config directory
# They are loaded at runtime by the launch file
Alpha = rospy.get_param("/moving_cube/alpha")
Epsilon = rospy.get_param("/moving_cube/epsilon")
Gamma = rospy.get_param("/moving_cube/gamma")
epsilon_discount = rospy.get_param("/moving_cube/epsilon_discount")
nepisodes = rospy.get_param("/moving_cube/nepisodes")
nsteps = rospy.get_param("/moving_cube/nsteps")
running_step = rospy.get_param("/moving_cube/running_step")
# Initialises the algorithm that we are going to use for learning
qlearn = qlearn.QLearn(actions=range(env.action_space.n),
alpha=Alpha, gamma=Gamma, epsilon=Epsilon)
initial_epsilon = qlearn.epsilon
start_time = time.time()
highest_reward = 0
# Starts the main training loop: the one about the episodes to do
for x in range(nepisodes):
rospy.logdebug("############### START EPISODE=>" + str(x))
cumulated_reward = 0
done = False
if qlearn.epsilon > 0.05:
qlearn.epsilon *= epsilon_discount
# Initialize the environment and get first state of the robot
observation = env.reset()
state = ''.join(map(str, observation))
episode_time = rospy.get_rostime().to_sec()
# for each episode, we test the robot for nsteps
for i in range(nsteps):
rospy.loginfo("############### Start Step=>"+str(i))
# Pick an action based on the current state
action = qlearn.chooseAction(state)
rospy.loginfo ("Next action is:%d", action)
# Execute the action in the environment and get feedback
observation, reward, done, info = env.step(action)
rospy.loginfo(str(observation) + " " + str(reward))
cumulated_reward += reward
if highest_reward < cumulated_reward:
highest_reward = cumulated_reward
nextState = ''.join(map(str, observation))
# Make the algorithm learn based on the results
rospy.logwarn("############### state we were=>" + str(state))
rospy.logwarn("############### action that we took=>" + str(action))
rospy.logwarn("############### reward that action gave=>" + str(reward))
rospy.logwarn("############### State in which we will start nect step=>" + str(nextState))
qlearn.learn(state, action, reward, nextState)
if not(done):
state = nextState
else:
rospy.loginfo ("DONE")
last_time_steps = numpy.append(last_time_steps, [int(i + 1)])
break
rospy.loginfo("############### END Step=>" + str(i))
#raw_input("Next Step...PRESS KEY")
#rospy.sleep(2.0)
m, s = divmod(int(time.time() - start_time), 60)
h, m = divmod(m, 60)
rospy.logwarn ( ("EP: "+str(x+1)+" - [alpha: "+str(round(qlearn.alpha,2))+" - gamma: "+str(round(qlearn.gamma,2))+" - epsilon: "+str(round(qlearn.epsilon,2))+"] - Reward: "+str(cumulated_reward)+" Time: %d:%02d:%02d" % (h, m, s)))
rospy.loginfo ( ("\n|"+str(nepisodes)+"|"+str(qlearn.alpha)+"|"+str(qlearn.gamma)+"|"+str(initial_epsilon)+"*"+str(epsilon_discount)+"|"+str(highest_reward)+"| PICTURE |"))
l = last_time_steps.tolist()
l.sort()
#print("Parameters: a="+str)
rospy.loginfo("Overall score: {:0.2f}".format(last_time_steps.mean()))
rospy.loginfo("Best 100 score: {:0.2f}".format(reduce(lambda x, y: x + y, l[-100:]) / len(l[-100:])))
env.close()
This the script for training.
From line 42 to line 49, we read the parameters for the q-learn algorithm from the parameter server which we’ll discuss later.
The training process is done from line 59 to line 108. For each step, the algorithm decides which action to take based on the current state, then it measures the observation, decides if the simulation is done and calculates rewards. It will keep learning to optimize the reward it gets.
The implementation of the q-learn algorithm was done by Victor Mayoral Vilches.
'''
Q-learning approach for different RL problems
as part of the basic series on reinforcement learning @
Inspired by https://gym.openai.com/evaluations/eval_kWknKOkPQ7izrixdhriurA
@author: Victor Mayoral Vilches <victor@erlerobotics.com>
'''
import random
class QLearn:
def __init__(self, actions, epsilon, alpha, gamma):
self.q = {}
self.epsilon = epsilon # exploration constant
self.alpha = alpha # discount constant
self.gamma = gamma # discount factor
self.actions = actions
def getQ(self, state, action):
return self.q.get((state, action), 0.0)
def learnQ(self, state, action, reward, value):
'''
Q-learning:
Q(s, a) += alpha * (reward(s,a) + max(Q(s') - Q(s,a))
'''
oldv = self.q.get((state, action), None)
if oldv is None:
self.q[(state, action)] = reward
else:
self.q[(state, action)] = oldv + self.alpha * (value - oldv)
def chooseAction(self, state, return_q=False):
q = [self.getQ(state, a) for a in self.actions]
maxQ = max(q)
if random.random() < self.epsilon:
minQ = min(q); mag = max(abs(minQ), abs(maxQ))
# add random values to all the actions, recalculate maxQ
q = [q[i] + random.random() * mag - .5 * mag for i in range(len(self.actions))]
maxQ = max(q)
count = q.count(maxQ)
# In case there're several state-action max values
# we select a random one among them
if count > 1:
best = [i for i in range(len(self.actions)) if q[i] == maxQ]
i = random.choice(best)
else:
i = q.index(maxQ)
action = self.actions[i]
if return_q: # if they want it, give it!
return action, q
return action
def learn(self, state1, action1, reward, state2):
maxqnew = max([self.getQ(state2, a) for a in self.actions])
self.learnQ(state1, action1, reward, reward + self.gamma*maxqnew)
We put the parameter for the q-learn separately in a file called one_disk_walk_openai_params.yaml under the config folder, so you can tweak the parameters much easier.
moving_cube: #namespace
running_step: 0.04 # amount of time the control will be executed
pos_step: 0.016 # increment in position for each command
#qlearn parameters
alpha: 0.1
gamma: 0.7
epsilon: 0.9
epsilon_discount: 0.999
nepisodes: 500
nsteps: 1000
number_splits: 10 #set to change the number of state splits for the continuous problem and also the number of env_variable splits
running_step: 0.06 # Time for each step
wait_time: 0.1 # Time to wait in the reset phases
n_actions: 5 # We have 3 actions
speed_step: 1.0 # Time to wait in the reset phases
init_roll_vel: 0.0 # Initial speed of the Roll Disk
roll_speed_fixed_value: 100.0 # Speed at which it will move forwards or backwards
roll_speed_increment_value: 10.0 # Increment that could be done in each step
max_distance: 2.0 # Maximum distance allowed for the RobotCube
max_pitch_angle: 0.2 # Maximum Angle radians in Pitch that we allow before terminating episode
max_yaw_angle: 0.1 # Maximum yaw angle deviation, after that it starts getting negative rewards
init_cube_pose:
x: 0.0
y: 0.0
z: 0.0
end_episode_points: 1000 # Points given when ending an episode
move_distance_reward_weight: 1000.0 # Multiplier for the moved distance reward, Ex: inc_d = 0.1 --> 100points
y_linear_speed_reward_weight: 1000.0 # Multiplier for moving fast in the y Axis
y_axis_angle_reward_weight: 1000.0 # Multiplier of angle of yaw, to keep it straight
It might take a lot of time to tune the parameters. In ROSDS, we offer the gym computer feature to help you run training with different parameters parallelly. If you are interested, please check our paid program.
If you want to learn more applications with OpenAI in ROS, please check our OpenAI course in the robot ignite academy.
Edit by Tony Huang.
[irp posts=”10198″ name=”ROS Projects OpenAI with Moving Cube Robot in Gazebo Step-by-Step #Part5″]
In this new ROS Project you are going to learn Step-by-Step how to create a moving cube and that it learns to move using OpenAI environment.
In this third video we continue with previous video setting up all the basics needed for moving the cube and getting sensor data for the reward and done functions
In the last post, we’ve briefly walked through the last part of cube_rl_utils.py which tests the robot. Today, let’s dive deeper into the code.
#!/usr/bin/env python
import time
import rospy
import math
import copy
import numpy
from std_msgs.msg import Float64
from sensor_msgs.msg import JointState
from nav_msgs.msg import Odometry
from geometry_msgs.msg import Point
from tf.transformations import euler_from_quaternion
class CubeRLUtils(object):
def __init__(self):
self.check_all_sensors_ready()
rospy.Subscriber("/moving_cube/joint_states", JointState, self.joints_callback)
rospy.Subscriber("/moving_cube/odom", Odometry, self.odom_callback)
self._roll_vel_pub = rospy.Publisher('/moving_cube/inertia_wheel_roll_joint_velocity_controller/command', Float64, queue_size=1)
self.check_publishers_connection()
def check_all_sensors_ready(self):
self.disk_joints_data = None
while self.disk_joints_data is None and not rospy.is_shutdown():
try:
self.disk_joints_data = rospy.wait_for_message("/moving_cube/joint_states", JointState, timeout=1.0)
rospy.loginfo("Current moving_cube/joint_states READY=>"+str(self.disk_joints_data))
except:
rospy.logerr("Current moving_cube/joint_states not ready yet, retrying for getting joint_states")
self.cube_odom_data = None
while self.disk_joints_data is None and not rospy.is_shutdown():
try:
self.cube_odom_data = rospy.wait_for_message("/moving_cube/odom", Odometry, timeout=1.0)
rospy.loginfo("Current /moving_cube/odom READY=>" + str(self.cube_odom_data))
except:
rospy.logerr("Current /moving_cube/odom not ready yet, retrying for getting odom")
rospy.loginfo("ALL SENSORS READY")
def check_publishers_connection(self):
"""
Checks that all the publishers are working
:return:
"""
rate = rospy.Rate(10) # 10hz
while (self._roll_vel_pub.get_num_connections() == 0 and not rospy.is_shutdown()):
rospy.loginfo("No susbribers to _roll_vel_pub yet so we wait and try again")
try:
rate.sleep()
except rospy.ROSInterruptException:
# This is to avoid error when world is rested, time when backwards.
pass
rospy.loginfo("_base_pub Publisher Connected")
rospy.loginfo("All Publishers READY")
def joints_callback(self, data):
self.joints = data
def odom_callback(self, data):
self.odom = data
...
The first part of the script is nothing special but creating publisher, subscriber and checking all the sensor.
# Reinforcement Learning Utility Code
def move_joints(self, roll_speed):
joint_speed_value = Float64()
joint_speed_value.data = roll_speed
rospy.loginfo("Single Disk Roll Velocity>>"+str(joint_speed_value))
self._roll_vel_pub.publish(joint_speed_value)
def get_cube_state(self):
# We convert from quaternions to euler
orientation_list = [self.odom.pose.pose.orientation.x,
self.odom.pose.pose.orientation.y,
self.odom.pose.pose.orientation.z,
self.odom.pose.pose.orientation.w]
roll, pitch, yaw = euler_from_quaternion(orientation_list)
# We get the distance from the origin
start_position = Point()
start_position.x = 0.0
start_position.y = 0.0
start_position.z = 0.0
distance = self.get_distance_from_point(start_position,
self.odom.pose.pose.position)
cube_state = [
round(self.joints.velocity[0],1),
round(distance,1),
round(roll,1),
round(pitch,1),
round(yaw,1)
]
return cube_state
def observation_checks(self, cube_state):
# MAximum distance to travel permited in meters from origin
max_distance=2.0
if (cube_state[1] > max_distance):
rospy.logerr("Cube Too Far==>"+str(cube_state[1]))
done = True
else:
rospy.loginfo("Cube NOT Too Far==>"+str(cube_state[1]))
done = False
return done
def get_distance_from_point(self, pstart, p_end):
"""
Given a Vector3 Object, get distance from current position
:param p_end:
:return:
"""
a = numpy.array((pstart.x, pstart.y, pstart.z))
b = numpy.array((p_end.x, p_end.y, p_end.z))
distance = numpy.linalg.norm(a - b)
return distance
def get_reward_for_observations(self, state):
# We reward it for lower speeds and distance traveled
speed = state[0]
distance = state[1]
# Positive Reinforcement
reward_distance = distance * 10.0
# Negative Reinforcement for magnitude of speed
reward_for_efective_movement = -1 * abs(speed)
reward = reward_distance + reward_for_efective_movement
rospy.loginfo("Reward_distance="+str(reward_distance))
rospy.loginfo("Reward_for_efective_movement= "+str(reward_for_efective_movement))
return reward
The second part is much more important. It prepares the elements for the reinforcement learning algorithm.
In the get_cube_state() function. We converted the sensor reading to cube state. We chose the joint velocity, distance, roll, pitch, yaw as the state. To get the roll, pitch, and yaw, we have to convert the odom from the quaternion to Euler angle.
We check if the simulation is done in the observation_checks() function and calculate reward in the get_reward_for_observations() function based on the distance the robot moved.
You can play with different parameters in the script to achieve a better reward. In the future, we will automate the learning process with the reinforcement learning algorithm.
Edit by: Tony Huang
[irp posts=”10079″ name=”ROS Projects OpenAI with Moving Cube Robot in Gazebo Step-by-Step #Part4″]
In this new ROS Project you are going to learn Step-by-Step how to create a moving cube and that it learns to move using OpenAI environment.
This second video is for learning the creation the basics of Reinforcement learning and how to connect to the various systems of the robot to get the state, perform actions and calculate rewards.
If you didn’t follow up, please check the link below for the last post.
[irp posts=”9744″ name=”[ROS Projects] OpenAI with Moving Cube Robot in Gazebo Step-by-Step Part1″]
Step 1. Clone the simulation
In order to make sure we have the same project. Please run the following command
cd ~/simulation_ws/src
git clone https://bitbucket.org/theconstructcore/moving_cube.git
NOTICE: please delete the previous code if you have problems to compile the code
In the /moving_cube/moving_cube_description/urdf/moving_cube.urdf file, please uncomment the following part. We’ll need this part to publish the odom topic.
To train the robot, let’s create a package for training under the catkin_ws/src
cd ~/catkin_ws/src
catkin_create_pkg my_moving_cube_traning_pkg rospy
Then we’ll create a script folder inside the package and put a file called cube_rl_utils.py inside it with the following content
#!/usr/bin/env python
import time
import rospy
import math
import copy
import numpy
from std_msgs.msg import Float64
from sensor_msgs.msg import JointState
from nav_msgs.msg import Odometry
from geometry_msgs.msg import Point
from tf.transformations import euler_from_quaternion
class CubeRLUtils(object):
def __init__(self):
self.check_all_sensors_ready()
rospy.Subscriber("/moving_cube/joint_states", JointState, self.joints_callback)
rospy.Subscriber("/moving_cube/odom", Odometry, self.odom_callback)
self._roll_vel_pub = rospy.Publisher('/moving_cube/inertia_wheel_roll_joint_velocity_controller/command', Float64, queue_size=1)
self.check_publishers_connection()
def check_all_sensors_ready(self):
self.disk_joints_data = None
while self.disk_joints_data is None and not rospy.is_shutdown():
try:
self.disk_joints_data = rospy.wait_for_message("/moving_cube/joint_states", JointState, timeout=1.0)
rospy.loginfo("Current moving_cube/joint_states READY=>"+str(self.disk_joints_data))
except:
rospy.logerr("Current moving_cube/joint_states not ready yet, retrying for getting joint_states")
self.cube_odom_data = None
while self.disk_joints_data is None and not rospy.is_shutdown():
try:
self.cube_odom_data = rospy.wait_for_message("/moving_cube/odom", Odometry, timeout=1.0)
rospy.loginfo("Current /moving_cube/odom READY=>" + str(self.cube_odom_data))
except:
rospy.logerr("Current /moving_cube/odom not ready yet, retrying for getting odom")
rospy.loginfo("ALL SENSORS READY")
def check_publishers_connection(self):
"""
Checks that all the publishers are working
:return:
"""
rate = rospy.Rate(10) # 10hz
while (self._roll_vel_pub.get_num_connections() == 0 and not rospy.is_shutdown()):
rospy.loginfo("No susbribers to _roll_vel_pub yet so we wait and try again")
try:
rate.sleep()
except rospy.ROSInterruptException:
# This is to avoid error when world is rested, time when backwards.
pass
rospy.loginfo("_base_pub Publisher Connected")
rospy.loginfo("All Publishers READY")
def joints_callback(self, data):
self.joints = data
def odom_callback(self, data):
self.odom = data
# Reinforcement Learning Utility Code
def move_joints(self, roll_speed):
joint_speed_value = Float64()
joint_speed_value.data = roll_speed
rospy.loginfo("Single Disk Roll Velocity>>"+str(joint_speed_value))
self._roll_vel_pub.publish(joint_speed_value)
def get_cube_state(self):
# We convert from quaternions to euler
orientation_list = [self.odom.pose.pose.orientation.x,
self.odom.pose.pose.orientation.y,
self.odom.pose.pose.orientation.z,
self.odom.pose.pose.orientation.w]
roll, pitch, yaw = euler_from_quaternion(orientation_list)
# We get the distance from the origin
start_position = Point()
start_position.x = 0.0
start_position.y = 0.0
start_position.z = 0.0
distance = self.get_distance_from_point(start_position,
self.odom.pose.pose.position)
cube_state = [
round(self.joints.velocity[0],1),
round(distance,1),
round(roll,1),
round(pitch,1),
round(yaw,1)
]
return cube_state
def observation_checks(self, cube_state):
# MAximum distance to travel permited in meters from origin
max_distance=2.0
if (cube_state[1] > max_distance):
rospy.logerr("Cube Too Far==>"+str(cube_state[1]))
done = True
else:
rospy.loginfo("Cube NOT Too Far==>"+str(cube_state[1]))
done = False
return done
def get_distance_from_point(self, pstart, p_end):
"""
Given a Vector3 Object, get distance from current position
:param p_end:
:return:
"""
a = numpy.array((pstart.x, pstart.y, pstart.z))
b = numpy.array((p_end.x, p_end.y, p_end.z))
distance = numpy.linalg.norm(a - b)
return distance
def get_reward_for_observations(self, state):
# We reward it for lower speeds and distance traveled
speed = state[0]
distance = state[1]
# Positive Reinforcement
reward_distance = distance * 10.0
# Negative Reinforcement for magnitude of speed
reward_for_efective_movement = -1 * abs(speed)
reward = reward_distance + reward_for_efective_movement
rospy.loginfo("Reward_distance="+str(reward_distance))
rospy.loginfo("Reward_for_efective_movement= "+str(reward_for_efective_movement))
return reward
def cube_rl_systems_test():
rospy.init_node('cube_rl_systems_test_node', anonymous=True, log_level=rospy.INFO)
cube_rl_utils_object = CubeRLUtils()
rospy.loginfo("Moving to Speed==>80")
cube_rl_utils_object.move_joints(roll_speed=80.0)
time.sleep(2)
rospy.loginfo("Moving to Speed==>-80")
cube_rl_utils_object.move_joints(roll_speed=-80.0)
time.sleep(2)
rospy.loginfo("Moving to Speed==>0.0")
cube_rl_utils_object.move_joints(roll_speed=0.0)
time.sleep(2)
cube_state = cube_rl_utils_object.get_cube_state()
done = cube_rl_utils_object.observation_checks(cube_state)
reward = cube_rl_utils_object.get_reward_for_observations(cube_state)
rospy.loginfo("Done==>"+str(done))
rospy.loginfo("Reward==>"+str(reward))
if __name__ == "__main__":
cube_rl_systems_test()
In this post, we’ll focus on the cube_rl_systems_test() function. The function uses the class to move the cube, get the observation, calculate reward and check if it’s done. To run it, you have to run the simulation first. Please go to Simulations->Select launch file-> main.launch
NOTICE: You have to unpause the simulation by clicking the arrow key in the simulation window
Then you can run the following command to run the script
cd ~/catkin_ws/src/my_moving_cube_training_pkg/script
chmod +x cube_rl_utils.py
cd ~/catkin_ws
source devel_setup.bash
rosrun my_moving_cube_training_pkg cube_rl_utils.py
You should see the cube moving around and the reward and the done state is calculated.
Edit by: Tony Huang
[irp posts=”9976″ name=”ROS Projects OpenAI with Moving Cube Robot in Gazebo Step-by-Step Part3″]
Step 1. Create a project in ROS Development Studio(ROSDS)
We’ll use ROSDS through this project in order to avoid setting up the environment, manage packages and etc. You can create a free account here if you haven’t had an account yet.
Step 2. Create package
Since this is a simulation, let’s create a package called my_moving_cube_description under the simulation_ws.
cd ~/simulation_ws/src
catkin_create_pkg my_moving_cube_description rospy
We’ll start by building the URDF description of the robot. To do that, we’ll create a new folder called urdf under the my_moving_cube_description directory and create a file called my_moving_cube.urdf inside it with the following initial content. The robot tag indicates the name of the robot – my_moving_cube.
<robot name="my_moving_cube">
...
</robot>
Then let’s create the first link inside the robot. This includes 3 parts :
inertial: It defines the physical property of the link. You can calculate the inertia of an object by using this tool: rosrun spawn_robot_tools_pkg inertial_calculator.py
collision: It defines the collision property when the object interacts with other objects in the simulation.
visual: It defines the visual property, how the object will visually show in the simulation.
You also need to define the material property after the link if you want to use it in gazebo. (NOTICE: the reference of the material property should have the same name as the link)
Then we create a moving_cube.yaml file under the my_moving_cube_description/config to define parameters for the controller
# .yaml config file
#
# The PID gains and controller settings must be saved in a yaml file that gets loaded
# to the param server via the roslaunch file (moving_cube_control.launch).
my_moving_cube:
# Publish all joint states -----------------------------------
# Creates the /joint_states topic necessary in ROS
joint_state_controller:
type: joint_state_controller/JointStateController
publish_rate: 30
# Effort Controllers ---------------------------------------
inertia_wheel_roll_joint_velocity_controller:
type: effort_controllers/JointVelocityController
joint: inertia_wheel_roll_joint
pid: {p: 1.0, i: 0.0, d: 0.0}
In the end, you should create one new launch file called moving_cube control.launch under the launch folder to launch the controller