def __init__(self, name, neighborDist=2.0, npoints=70):
print "Starting.", name
self.name = name
self.npoints = npoints
self.neighborDist = neighborDist
# This is a list of the robots on our team, assumed to contain
# Neighbor() objects, as above.
self.neighbors = []
# This is a list of the robots within neighborDist of us.
self.neighborNames = []
# This is an array of points in configuration space indicating
# where we need to go next.
self.deltaConfig = np.array([[]])
self.twist = Twist()
self.maxSpeed = 0.5 # 0.25
self.maxTurn = 1.0 # .6
self.kSpeed = 15.0
self.kTurn = 13.0
# These are coefficients for the opinion formation. Tune and forget.
self.aw = 0.5
self.sp = 0.2 # Without this term, you get an exploding ring more often.
self.av = 0.6
self.se = 0.0
self.maxDist = 2.5
self.phi = Phi(self.name, self.aw, self.sp, self.av, self.se,
self.maxDist)
self.movePeriod = 5.0
self.nextMoveTime = rospy.Time.now().to_sec() + self.movePeriod
# Estimated variance in the position measurements.
self.std = 0.01
# When rating a neighbor's opinion about our desired position,
# use this many bins. Shouldn't be a radically different
# order of magnitude from npoints.
self.nbins = (5, 5)
rospy.sleep(1.0)
self.odomSub = rospy.Subscriber(self.name + "/odom",
Odometry,
self.trackOdom,
queue_size=1,
buff_size=250)
self.advice = rospy.Service(self.name + "/advice", MRFQuery,
self.takeAdvice)
self.busy = False
def __init__(self, name, neighborDist = 2.0, npoints = 70):
print "Starting." , name
self.name = name
self.npoints = npoints
self.neighborDist = neighborDist
# This is a list of the robots on our team, assumed to contain
# Neighbor() objects, as above.
self.neighbors = []
# This is a list of the robots within neighborDist of us.
self.neighborNames = []
# This is an array of points in configuration space indicating
# where we need to go next.
self.deltaConfig = np.array([[]])
self.twist = Twist()
self.maxSpeed = 0.5 # 0.25
self.maxTurn = 1.0 # .6
self.kSpeed = 15.0
self.kTurn = 13.0
# These are coefficients for the opinion formation. Tune and forget.
self.aw = 0.5
self.sp = 0.2 # Without this term, you get an exploding ring more often.
self.av = 0.6
self.se = 0.0
self.maxDist = 2.5
self.phi = Phi(self.name,
self.aw, self.sp, self.av, self.se,
self.maxDist)
self.movePeriod = 5.0
self.nextMoveTime = rospy.Time.now().to_sec() + self.movePeriod
# Estimated variance in the position measurements.
self.std = 0.01
# When rating a neighbor's opinion about our desired position,
# use this many bins. Shouldn't be a radically different
# order of magnitude from npoints.
self.nbins = (5,5)
rospy.sleep(1.0)
self.odomSub = rospy.Subscriber(self.name + "/odom", Odometry,
self.trackOdom,
queue_size = 1, buff_size = 250)
self.advice = rospy.Service(self.name + "/advice", MRFQuery,
self.takeAdvice)
self.busy = False
def __init__(self, name, npoints = 7, away = 1.3, space = 0.6,
avoid = 0.5, seek = 4.0, maxDist = 2.0,
velPub = False, goalPub = False):
print "Starting:" , name
self.name = name
self.pos = Belief(data_type = "position", data_sub_type = self.name,
pers = (0.0, 0.0, 2*math.pi))
self.npoints = npoints
# These are coefficients for the opinion formation.
# Away controls the desire to move apart.
self.away = away
# Without this space term, you get an exploding ring more often with
# nothing in the middle.
self.space = space
# Controls the predilection to avoid previously-encountered obstacles.
self.avoid = avoid
# Controls interest in new territory.
self.seek = seek
# Poorly named -- the potential function in the phi.away() method
# attempts to put this distance between each node.
self.maxDist = maxDist
# This is a list of neighboring robots, assumed to contain
# Neighbor() objects, as defined above.
self.neighbors = []
# This is an array of points in configuration space indicating
# where we need to go next.
self.deltaConfig = np.array([[]])
self.twist = Twist()
#** These are moved to gnav
self.maxSpeed = 0.25
self.maxTurn = 0.35
self.kSpeed = 1.0
self.kTurn = 1.0
self.phi = Phi(self.name,
self.away, self.space, self.avoid, self.seek,
self.maxDist)
self.movePeriod = 3.0
self.nextMoveTime = rospy.Time.now().to_sec() + self.movePeriod
# When this is in the future, we are still dealing with a bump. Don't
# register other obstacles within a couple of seconds.
self.bumpFreeTime = 0.0
# Estimated variance in the position measurements. This is important
# in the resampling of the guesses.
self.std = 0.1
# When rating a neighbor's opinion about our desired position,
# use this many bins. Shouldn't be a radically different
# order of magnitude from npoints.
self.nbins = (5,5)
self.velPub = velPub
self.goalPub = goalPub
try:
# assert velPub
assert goalPub
except:
print "must specify a goal publisher"
self.busy = False
class CoverageControl(object):
def __init__(self, name, npoints = 7, away = 1.3, space = 0.6,
avoid = 0.5, seek = 4.0, maxDist = 2.0,
velPub = False, goalPub = False):
print "Starting:" , name
self.name = name
self.pos = Belief(data_type = "position", data_sub_type = self.name,
pers = (0.0, 0.0, 2*math.pi))
self.npoints = npoints
# These are coefficients for the opinion formation.
# Away controls the desire to move apart.
self.away = away
# Without this space term, you get an exploding ring more often with
# nothing in the middle.
self.space = space
# Controls the predilection to avoid previously-encountered obstacles.
self.avoid = avoid
# Controls interest in new territory.
self.seek = seek
# Poorly named -- the potential function in the phi.away() method
# attempts to put this distance between each node.
self.maxDist = maxDist
# This is a list of neighboring robots, assumed to contain
# Neighbor() objects, as defined above.
self.neighbors = []
# This is an array of points in configuration space indicating
# where we need to go next.
self.deltaConfig = np.array([[]])
self.twist = Twist()
#** These are moved to gnav
self.maxSpeed = 0.25
self.maxTurn = 0.35
self.kSpeed = 1.0
self.kTurn = 1.0
self.phi = Phi(self.name,
self.away, self.space, self.avoid, self.seek,
self.maxDist)
self.movePeriod = 3.0
self.nextMoveTime = rospy.Time.now().to_sec() + self.movePeriod
# When this is in the future, we are still dealing with a bump. Don't
# register other obstacles within a couple of seconds.
self.bumpFreeTime = 0.0
# Estimated variance in the position measurements. This is important
# in the resampling of the guesses.
self.std = 0.1
# When rating a neighbor's opinion about our desired position,
# use this many bins. Shouldn't be a radically different
# order of magnitude from npoints.
self.nbins = (5,5)
self.velPub = velPub
self.goalPub = goalPub
try:
# assert velPub
assert goalPub
except:
print "must specify a goal publisher"
self.busy = False
def recordNeighbors(self, req):
"""
At any time, the self.neighbors list should contain a Neighbor
object for each robot whose position we can measure directly.
We are receiving a list of robots and distances (signal
strengths) via the /olsr/neighbors topic and accumulating.
Use this as a callback to the /olsr/neighbors topic. The
result of calling this is an updated self.neighbors array.
"""
### This is a hack until the OLSR ROS node is working. We just
### use a fixed list of all the robots and subtract our name.
rm = Room(sourceName = self.name)
rm.names = req
rm.locations = [ Pt(point = (1.0, 0.0)), Pt(point = (1.0, 0.0)) ]
rm.names.remove(self.name)
self.neighbors = []
for name in rm.names:
self.neighbors.append(Neighbor(name = name))
def trackSensors(self, sensors):
if sensors.bumpLeft or sensors.bumpRight:
now = rospy.Time.now().to_sec()
if now > self.bumpFreeTime:
# Register the obstacle with the phi function.
obstaclePos = (self.pos.means[0]+0.3*math.cos(self.pos.means[2]),
self.pos.means[1]+0.3*math.sin(self.pos.means[2]),
0.0 )
self.phi.addObstacle(obstaclePos)
# Set bump time (when it's ok to pay attention to the phi again)
self.bumpFreeTime = rospy.Time.now().to_sec() + 1.0
def trackPos(self, req):
"""
This gets called as often as a new position estimate is
issued by the perfesser.
"""
if rospy.Time.now().to_sec() > self.bumpFreeTime:
self.bumpFreeTime = 0.0
self.pos = req
# We are relying on the Belief.means to have been filled in by
# the sending process. Should check.
assert sum(self.pos.means) != 0.0
self.phi.updateOccupancy(self.pos.means)
now = rospy.Time.now().to_sec()
if self.nextMoveTime - now < self.movePeriod:
self.evaluateMove()
self.nextMoveTime = now + self.movePeriod
def takeAdvice(self, req):
"""
Takes a piece of advice from its neighbor, rates it, and
returns the rating. The advice is a distribution of points to
which the other robot thinks this robot ought to go. We
compare that advice to our own intentions and return ratings
indicating the degree of agreement between the two: 1.0 is
complete agreement and 0.0 is complete disagreement. We never
return exactly 0.0 as a matter of policy.
"""
print "entering takeAdvice:", str(req)
if not req.question:
return MRFQueryResponse(no_data = True)
if not self.deltaConfig.any():
return MRFQueryResponse(no_data = True)
# Allocate our intention into bins.
question = self.ptsToArray(req.question)
# Establish bins so that there is one point in each outer bin.
# (The Histogram class in perfesser is set up so the outer
# bins will be empty, which is why we're not using it here.
mins = [ min(self.deltaConfig[:,0]), min(self.deltaConfig[:,1]) ]
maxs = [ max(self.deltaConfig[:,0]), max(self.deltaConfig[:,1]) ]
bins = [ np.linspace(mn + 1.e-10, mx - 1.e-10, bn - 1) for \
mn, mx, bn in zip(mins, maxs, self.nbins) ]
hist = np.zeros(self.nbins)
# Create a distribution of points that includes neighbors'
# opinions, omitting the one asking the question.
phi = self.deltaConfig[:,(0,1)]
messageProduct = [ 1.0 ] * phi.shape[0]
for nb in self.neighbors:
if (nb.name != req.asker) and (nb.lastReply.points):
messageProduct = [ m * p.point[0] for m,p in \
zip(messageProduct,
nb.lastReply.points) ]
phi = self.resample(phi, messageProduct)
# Populate bins
for delta in phi:
hist[sum(delta[0] > bins[0])][sum(delta[1] > bins[1])] += 1.0
n = sum(sum(hist))
# Formulate a reply by binning the input data.
m = MRFQueryResponse(no_data = False,
source_stamp = req.source_stamp)
for q in req.question:
m.reply.append(Pt(point=( \
hist[sum(q.point[0] > bins[0])][sum(q.point[1] > bins[1])]/n, )))
# print "entering service: ", self.name
# print self.name, "'s opinion: ", self.deltaConfig
# print "question: ",question
# print "hist: ", hist
# print "bins: ", bins
# print m
return m
def ptsToArray(self, pts):
return np.array([p.point for p in pts])
def arrayToPts(self, array):
return [ Pt(point = a) for a in array ]
def setAdvice(self, name, advice):
"""
Takes a name of a robot and an array of advice and stores it
in the neighbor array -- if there is an appropriate entry for
it there.
"""
# print "setAdvice:", name, advice
for nb in self.neighbors:
if nb.name == name:
nb.lastQuestion.points = self.arrayToPts(advice)
def evaluateMove(self):
"""
Formulates a distribution of points representing a belief
about the direction in which we should move, along with our
neighbors. The belief is based on our measurements of our own
location, as well as the location of the other robots, along
with some goals (i.e. spread out, explore, etc).
To be explicit, inputs are these:
- Location of self
- Location of neighbors
- Suggestions from neighbors about where self should move.
- Replies by neighbors to suggestions about where they
should move.
The first two form the input to the phi() function. The third
is the input to the psi() function, and the fourth is the
messageProduct referenced below.
Definition: "configuration space" is the n-dimensional space
where each point represents a particular configuration of all
the robots in it. So a one-dimensional room containing three
robots corresponds to a three-dimensional configuration
space. A two-dimensional room with five robots makes ten
dimensions in configuration space. Seems confusing, but it
makes some of the math easier.
"""
if self.busy or (self.bumpFreeTime > 0.0):
return
self.busy = True
print "processing... (%s)" % (self.name,)
# Marshal the inputs: our position and the positions of our
# neighbors. Park them all in an array where each row
# represents a position in configuration space and the
# collection of rows represent the distribution of those
# points.
currentConfig = np.array(self.ptsToArray(self.pos.points))[:,(0,1)]
# For 2d this should be two entries with X and Y
self.deltaNames = [ self.name + "/dx", self.name + "/dy" ]
print "NEIGHBORS>>>", [nb.name for nb in self.neighbors], "<<<"
for nb in self.neighbors:
if nb.location.points:
currentConfig = \
np.hstack([currentConfig,
self.ptsToArray(nb.location.points)[:,(0,1)]])
self.deltaNames += [ nb.name + "/dx", nb.name + "/dy" ]
# The currentConfig is now constructed.
print "got deltaNames: ", self.deltaNames
print "cc[0]:", currentConfig[0]
print "cc[1]:", currentConfig[1]
print "cc[2]:", currentConfig[2]
#print str(currentConfig)
# Run through the points in the current configuration and
# generate predictions for each point in the distribution.
deltaPts = []
for cpoint in currentConfig:
deltaPts.append(self.phi.phi(cpoint))
self.deltaConfig = np.array(deltaPts)
# Break up the array, and store the opinions about each
# neighbor in the neighbor list. We do it this way in case
# the neighbor list has changed since the marshaling.
machineNames = [ self.deltaNames[i].split("/")[0] \
for i in range(0, len(self.deltaNames), 2) ]
print "mnames:", machineNames
for nm,dc in zip(machineNames,
np.hsplit(self.deltaConfig,
self.deltaConfig.shape[1]/2)):
# Record the advice given to that robot.
# print "setting advice", nm, dc
self.setAdvice(nm, dc)
#print str(self.deltaConfig)
# Notify each neighbor of our opinion (and record their
# responses).
messageProduct = [ 1.0 ] * self.npoints
for nb in self.neighbors:
print "about to advise:", nb.name
# If we don't already have an opinion about where this
# robot should be then pass for the moment and we'll get
# it the next time around.
if not nb.lastQuestion.points:
break
m = MRFQueryRequest(source_stamp = rospy.Time.now(),
asker = self.name,
asked = nb.name,
question = nb.lastQuestion.points)
print "advising: ", nb.name
r = nb.advise(m)
if r.reply:
nb.lastReply.points = r.reply
nb.lastReply.source_stamp = r.source_stamp
messageProduct = [ lr.point[0] * mp \
for lr, mp in zip(r.reply,
messageProduct) ]
# Resample based on the neighbor's responses.
desiredConfig = self.resample(self.deltaConfig, messageProduct)
# The first set of coordinates in this is the desired place
# for us to move.
moves = desiredConfig[:,(0,1)]
self.goThatWay(moves)
self.busy = False
def displayConf(self, name, configArray):
"""
Print the configuration array, or rather a mean of it.
"""
out = []
for row in configArray.transpose():
out.append(np.mean(row))
fmt = name + " " + self.name + ": (" + "%.3f," * len(out) + ")"
return fmt % tuple(out)
def goThatWay(self, moves):
"""
Takes an array of directions to move in and selects a point in
space in that direction. The choice of point is fairly
arbitrary. We try to multiply the direction by a factor large
enough that it won't be within the gnav.zeroDistance of the
current location, unless the intent is really that we should
not move. (That's what the kGo factor is for.)
"""
kGo = 2.0
print "moves", moves
meanMove = [ np.mean(moves[:, 0]), np.mean(moves[:, 1]) ]
sumsq = sum([d**2 for d in meanMove])**0.5
# If the target point is close, don't normalize. Otherwise, set
# a goal in the general direction
if sumsq > 1.0:
meanMove = [ kGo * x/sumsq for x in meanMove ]
self.goal = (self.pos.means[0] + meanMove[0],
self.pos.means[1] + meanMove[1], 0.0)
print "We're at: (%.3f,%.3f)" % (self.pos.means[0], self.pos.means[1])
print "headed to: (%.3f,%.3f)" % (self.goal[0], self.goal[1])
# Eventually this should publish to a goal-setting topic, to be
# read by some goal-setter over at gnav.
self.goalPub.publish(Pt(point=self.goal))
def stepThatWay(self):
"""
Not used.
"""
fX = self.goal[0] - self.pos.means[0]
fY = self.goal[1] - self.pos.means[1]
fM = math.hypot(fX, fY)
fD = math.atan2(fY, fX)
# Convert to robot frame
fDR = fD - self.pos.means[2]
if math.fabs(fDR) > math.pi:
fDR -= math.copysign(2 * math.pi, fDR)
if math.fabs(fDR) > math.pi/10.0:
provX = 0.0
else:
provX = min(self.maxSpeed, self.kSpeed * fM)
provZ = math.fabs(self.kTurn * fDR)
provZ = min(self.maxTurn, provZ)
provZ = math.copysign(provZ, fDR)
twist = Twist()
twist.linear.x = provX
twist.angular.z = provZ
return twist
def weightedSample(self, pointProbs, numSamples):
"""
The following sampling algorithm is adapted from a sneaky
algorithm for coming up with a selection of N random numbers
between 0 and Q. The idea is that the probability that all N
of the numbers are less than some X is P=(X/Q)**N. Solve for
X in terms of P, and you get an equation that will pick the
largest number in that distribution for some random P. Do
that N times, and you get N samples.
The input is an indexed array of probabilities (use
enumerate(P) for the input, e.g. and the number of samples to
take from that array. The output is an iterator that will
provide the indices you can use to sample the original array
with, as in:
output = locArray[weightedSample(enumerate(locProbs), n)
This won't exactly work, since the enumerate has to be turned
into a sequence before it can be subscripted (as below) and
the same is true of the function output.
For more: stackoverflow.com/questions/2140787
"""
total = sum(prob for i, prob in pointProbs)
assert(not math.isnan(total))
j = 0
i, prob = pointProbs[j]
while numSamples:
x = total * (1 - np.random.random() ** (1.0/numSamples))
total -= x
while x > prob:
x -= prob
j += 1
i, prob = pointProbs[j]
prob -= x
yield i
numSamples -= 1
def resample(self, array, ratings):
"""
Resamples the distribution represented by the samples in
array, based on the list of ratings. There should be one
rating for each row of the array, and each rating should be
greater than zero and less than or equal to 1.0. The ratings
list need not be normed.
"""
#print "resample according to: ", ratings
newseq = self.weightedSample([(i,p) for i,p in enumerate(ratings)],
len(ratings))
#print "comes up with: ", [p for p in newseq]
wiggle = lambda x: self.std * (np.random.random() - 0.5) + x
out = []
for pt in array[ [ i for i in newseq ], :]:
out.append([ wiggle(pt[0]), wiggle(pt[1]) ])
return np.array(out)
class coverControl(object):
def __init__(self, name, neighborDist=2.0, npoints=70):
print "Starting.", name
self.name = name
self.npoints = npoints
self.neighborDist = neighborDist
# This is a list of the robots on our team, assumed to contain
# Neighbor() objects, as above.
self.neighbors = []
# This is a list of the robots within neighborDist of us.
self.neighborNames = []
# This is an array of points in configuration space indicating
# where we need to go next.
self.deltaConfig = np.array([[]])
self.twist = Twist()
self.maxSpeed = 0.5 # 0.25
self.maxTurn = 1.0 # .6
self.kSpeed = 15.0
self.kTurn = 13.0
# These are coefficients for the opinion formation. Tune and forget.
self.aw = 0.5
self.sp = 0.2 # Without this term, you get an exploding ring more often.
self.av = 0.6
self.se = 0.0
self.maxDist = 2.5
self.phi = Phi(self.name, self.aw, self.sp, self.av, self.se,
self.maxDist)
self.movePeriod = 5.0
self.nextMoveTime = rospy.Time.now().to_sec() + self.movePeriod
# Estimated variance in the position measurements.
self.std = 0.01
# When rating a neighbor's opinion about our desired position,
# use this many bins. Shouldn't be a radically different
# order of magnitude from npoints.
self.nbins = (5, 5)
rospy.sleep(1.0)
self.odomSub = rospy.Subscriber(self.name + "/odom",
Odometry,
self.trackOdom,
queue_size=1,
buff_size=250)
self.advice = rospy.Service(self.name + "/advice", MRFQuery,
self.takeAdvice)
self.busy = False
def addNeighbor(self, name):
self.neighbors.append(Neighbor(name))
def addNeighbors(self, names):
for name in names:
self.addNeighbor(name)
def recordNeighbors(self):
"""
At any time, the self.neighbors list should contain an entry for
all the robots in the experiment and the self.nbrNames list should
contain the names of the robots whose positions are close enough to
consider true neighbors.
"""
self.neighborNames = []
for nb in self.neighbors:
if self.dist(self.pos, nb.location.means) < self.neighborDist:
self.neighborNames.append(nb.name)
if not self.neighborNames:
# No neighbors -- should act on that. For now just
# return.
return
def trackBump(self, req):
self.phi.addObstacle(req.means)
def trackPos(self, req):
"""
This gets called as often as a new position estimate is issued.
"""
self.posBelief = req
self.pos = req.means
self.phi.updateOccupancy(self.pos)
now = rospy.Time.now().to_sec()
if self.nextMoveTime - now < self.movePeriod:
self.evaluateMove()
self.nextMoveTime = now + self.movePeriod
def takeAdvice(self, req):
"""
Takes a piece of advice from its neighbor, rates it, and
returns the rating. The advice is a distribution of points to
which the other robot thinks this robot ought to go. We
compare that advice to our own intentions and return ratings
indicating the degree of agreement between the two: 1.0 is
complete agreement and 0.0 is complete disagreement. We never
return exactly 0.0 as a matter of policy.
"""
if not req.question:
return MRFQueryResponse(no_data=True)
if not self.deltaConfig.any():
return MRFQueryResponse(no_data=True)
# Allocate our intention into bins.
question = self.ptsToArray(req.question)
# Establish bins so that there is one point in each outer bin.
# (The Histogram class in perfesser is set up so the outer
# bins will be empty, which is why we're not using it here.
mins = [min(self.deltaConfig[:, 0]), min(self.deltaConfig[:, 1])]
maxs = [max(self.deltaConfig[:, 0]), max(self.deltaConfig[:, 1])]
bins = [ np.linspace(mn + 1.e-10, mx - 1.e-10, bn) for \
mn,mx,bin in zip(mins, maxs, self.nbins) ]
hist = np.zeros(self.nbins)
# Create a distribution of points that includes neighbor's
# opinions except for the one asking the question.
phi = self.deltaConfig[:, (0, 1)]
messageProduct = [1.0] * phi.shape[0]
for nb in self.neighbors:
if (nb.name != req.asker) and (nb.lastReply.points):
messageProduct = [ m * p.point[0] for m,p in \
zip(messageProduct,
nb.lastReply.points) ]
phi = self.resample(phi, messageProduct)
# Populate bins
for delta in phi:
hist[sum(delta[0] > bins[0])][sum(delta[1] > bins[1])] += 1.0
n = sum(sum(hist))
# Formulate a reply by binning the input data.
m = MRFQueryResponse(no_data=False, source_stamp=req.source_stamp)
for q in req.question:
m.reply.append(Pt(point=( \
hist[sum(q.point[0] > bins[0])][sum(q.point[1] > bins[1])]/n, )))
# print "entering service: ", self.name
# print self.name, "'s opinion: ", self.deltaConfig
# print "question: ",question
# print "hist: ", hist
# print "bins: ", bins
# print m
return m
def ptsToArray(self, pts):
return np.array([p.point for p in pts])
def arrayToPts(self, array):
return [Pt(point=a) for a in array]
def setAdvice(self, name, advice):
"""
Takes a name of a robot and an array of advice and stores it
in the neighbor array -- if appropriate.
"""
for nb in self.neighbors:
if nb.name == name:
nb.lastQuestion.points = self.arrayToPts(advice)
def evaluateMove(self):
"""
Formulates a distribution of points representing a belief
about the direction in which we should move, along with our
neighbors. The belief is based on our measurements of our own
location, as well as the location of the other robots, along
with some goals (i.e. spread out, explore, etc).
To be explicit, inputs are these:
- Location of self
- Location of neighbors
- Suggestions from neighbors about where self should move.
- Replies by neighbors to suggestions about where they
should move.
The first two form the input to the phi() function. The third
is the input to the psi() function, and the fourth is the
messageProduct referenced below.
Definition: "configuration space" is the n-dimensional space
where each point represents a particular configuration of all
the robots in it. So a one-dimensional room containing three
robots corresponds to a three-dimensional configuration
space. A two-dimensional room with five robots makes ten
dimensions in configuration space. Seems confusing, but it
makes some of the math easier.
"""
# Marshal the inputs: our position and the positions of our
# neighbors. Park them all in an array where each row
# represents a position in configuration space and the
# collection of rows represent the distribution of those
# points.
currentConfig = np.array(self.ptsToArray(self.posBelief.points))[:,
(0,
1)]
# For 2d this should be two entries with X and Y
self.deltaNames = [self.name + "/dx", self.name + "/dy"]
# All the robots are in the neighbors list, but only the nearby ones
# are to be used.
for nb in self.neighbors:
if nb.location.points:
currentConfig = np.hstack(
[currentConfig,
self.ptsToArray(nb.location.points)])
self.deltaNames += [nb.name + "/dx", nb.name + "/dy"]
# Run through the points in the current configuration and
# generate predictions for each point in the distribution.
deltaPts = []
for cpoint in currentConfig:
deltaPts.append(self.phi.phi(cpoint))
self.deltaConfig = np.array(deltaPts)
# Break up the array, and store the opinions about each
# neighbor in the neighbor list. We do it this way in case
# the neighbor list has changed since the marshaling.
for nm, dc in zip(
self.deltaNames,
np.hsplit(self.deltaConfig, self.deltaConfig.shape[1])):
self.setAdvice(nm, dc)
# Notify each neighbor of our opinion (and record their
# responses).
messageProduct = [1.0] * self.npoints
for nb in self.neighbors:
# If we don't already have an opinion about where this
# robot should be then pass for the moment and we'll get
# it the next time around.
if not nb.lastQuestion.points:
break
m = MRFQueryRequest(source_stamp=rospy.Time.now(),
asker=self.name,
asked=nb.name,
question=nb.lastQuestion.points)
r = nb.advise(m)
if r.reply:
nb.lastReply.points = r.reply
nb.lastReply.source_stamp = r.source_stamp
messageProduct = [ lr.point[0] * mp \
for lr, mp in zip(r.reply,
messageProduct) ]
# Resample based on the neighbor's responses.
desiredConfig = self.resample(self.deltaConfig, messageProduct)
# The first set of coordinates in this is the desired place
# for us to move.
moves = desiredConfig[:, (0, 1)]
self.goThatWay(moves)
self.busy = False
def goThatWay(self, moves):
"""
Takes an array of points to move to and selects a direction in
which to move in order to get us there.
"""
self.goal = (self.pos[0] + np.mean(moves[:, 0]),
self.pos[1] + np.mean(moves[:, 1]), 0.0)
####### SET NEW GOAL HERE #######
def weightedSample(self, pointProbs, numSamples):
"""
The following sampling algorithm is adapted from a sneaky
algorithm for coming up with a selection of N random numbers
between 0 and Q. The idea is that the probability that all N
of the numbers are less than some X is P=(X/Q)**N. Solve for
X in terms of P, and you get an equation that will pick the
largest number in that distribution for some random P. Do
that N times, and you get N samples.
The input is an indexed array of probabilities (use
enumerate(P) for the input, e.g. and the number of samples to
take from that array. The output is an iterator that will
provide the indices you can use to sample the original array
with, as in:
output = locArray[weightedSample(enumerate(locProbs), n)
This won't exactly work, since the enumerate has to be turned
into a sequence before it can be subscripted (as below) and
the same is true of the function output.
For more: stackoverflow.com/questions/2140787
"""
total = sum(prob for i, prob in pointProbs)
assert (not math.isnan(total))
j = 0
i, prob = pointProbs[j]
while numSamples:
x = total * (1 - np.random.random()**(1.0 / numSamples))
total -= x
while x > prob:
x -= prob
j += 1
i, prob = pointProbs[j]
prob -= x
yield i
numSamples -= 1
def resample(self, array, ratings):
"""
Resamples the distribution represented by the samples in
array, based on the list of ratings. There should be one
rating for each row of the array, and each rating should be
greater than zero and less than or equal to 1.0. The ratings
list need not be normed.
"""
#print "resample according to: ", ratings
newseq = self.weightedSample([(i, p) for i, p in enumerate(ratings)],
len(ratings))
#print "comes up with: ", [p for p in newseq]
wiggle = lambda x: self.std * (np.random.random() - 0.5) + x
out = []
for pt in array[[i for i in newseq], :]:
out.append([wiggle(pt[0]), wiggle(pt[1])])
return np.array(out)
def __init__(self, name, initpos=(0.0, 0.0, 0.0), npoints=10):
print "Starting.", name
self.name = name
self.pos = initpos
self.npoints = npoints
# This is a list of neighboring robots, assumed to contain
# Neighbor() objects, as above.
self.neighbors = []
self.velPub = rospy.Publisher(self.name + "/cmd_vel", Twist)
self.posPub = rospy.Publisher(self.name + "/position", Belief)
# This is an array of points in configuration space indicating
# where we need to go next.
self.deltaConfig = np.array([[]])
self.twist = Twist()
self.maxSpeed = 0.5 # 0.25
self.maxTurn = 1.0 # .6
self.kSpeed = 15.0
self.kTurn = 13.0
# These are coefficients for the opinion formation.
self.aw = 13.
self.sp = 0.1 # Without this term, you get an exploding ring more often.
self.av = 3.
self.se = 250.0
self.maxDist = 2.5
self.phi = Phi(self.name, self.aw, self.sp, self.av, self.se,
self.maxDist)
self.movePeriod = 5.0
self.nextMoveTime = rospy.Time.now().to_sec() + self.movePeriod
# When this is in the future, we are still dealing with a bump.
self.bumpFreeTime = 0.0
# Estimated variance in the position measurements.
self.std = 0.01
# When rating a neighbor's opinion about our desired position,
# use this many bins. Shouldn't be a radically different
# order of magnitude from npoints.
self.nbins = (5, 5)
rospy.sleep(1.0)
self.sensorSub = rospy.Subscriber(self.name + "/sensors", SensorPacket,
self.trackSensors)
self.odomSub = rospy.Subscriber(self.name + "/odom",
Odometry,
self.trackOdom,
queue_size=1,
buff_size=250)
self.advice = rospy.Service(self.name + "/advice", MRFQuery,
self.takeAdvice)
self.roomLoc = rospy.Subscriber("/room/locations", Room2D,
self.recordNeighbors)
self.busy = False
class modelControl(object):
def __init__(self, name, initpos=(0.0, 0.0, 0.0), npoints=10):
print "Starting.", name
self.name = name
self.pos = initpos
self.npoints = npoints
# This is a list of neighboring robots, assumed to contain
# Neighbor() objects, as above.
self.neighbors = []
self.velPub = rospy.Publisher(self.name + "/cmd_vel", Twist)
self.posPub = rospy.Publisher(self.name + "/position", Belief)
# This is an array of points in configuration space indicating
# where we need to go next.
self.deltaConfig = np.array([[]])
self.twist = Twist()
self.maxSpeed = 0.5 # 0.25
self.maxTurn = 1.0 # .6
self.kSpeed = 15.0
self.kTurn = 13.0
# These are coefficients for the opinion formation.
self.aw = 13.
self.sp = 0.1 # Without this term, you get an exploding ring more often.
self.av = 3.
self.se = 250.0
self.maxDist = 2.5
self.phi = Phi(self.name, self.aw, self.sp, self.av, self.se,
self.maxDist)
self.movePeriod = 5.0
self.nextMoveTime = rospy.Time.now().to_sec() + self.movePeriod
# When this is in the future, we are still dealing with a bump.
self.bumpFreeTime = 0.0
# Estimated variance in the position measurements.
self.std = 0.01
# When rating a neighbor's opinion about our desired position,
# use this many bins. Shouldn't be a radically different
# order of magnitude from npoints.
self.nbins = (5, 5)
rospy.sleep(1.0)
self.sensorSub = rospy.Subscriber(self.name + "/sensors", SensorPacket,
self.trackSensors)
self.odomSub = rospy.Subscriber(self.name + "/odom",
Odometry,
self.trackOdom,
queue_size=1,
buff_size=250)
self.advice = rospy.Service(self.name + "/advice", MRFQuery,
self.takeAdvice)
self.roomLoc = rospy.Subscriber("/room/locations", Room2D,
self.recordNeighbors)
self.busy = False
def recordNeighbors(self, req):
"""
At any time, the self.neighbors list should contain a Neighbor
object for each robot whose position we can measure directly.
For the purpose of simulation, we are receiving a list of
robots and positions via the /room/locations topic and
accumulating. Use this as a callback to the /room/locations
topic. The result of calling this is an updated
self.neighbors array.
"""
distances = []
nabe = False
for nb in req.robots:
if nb.sourceName == self.name:
nabe = nb
break
if not nabe:
# No neighbors -- should act on that. For now just
# return.
return
# Now we have a list of robots and locations that are neighbors
# Check them against the existing list and edit accordingly.
nameList = [nb.name for nb in self.neighbors]
for name in nabe.names:
if not (name in nameList):
self.neighbors.append(Neighbor(name=name))
for nb in self.neighbors:
if not (nb.name in nabe.names):
self.neighbors.remove(nb)
def getPosEst(self):
meanloc = (np.random.normal(loc=self.pos[0],
scale=self.std / 2.,
size=1)[0],
np.random.normal(loc=self.pos[1],
scale=self.std / 2.,
size=1)[0])
out = Belief(source_stamp=rospy.Time.now(),
source_data="estimate",
data_type="location",
data_sub_type=self.name,
sender=self.name,
dim_names=[self.name + "/xpos", self.name + "/ypos"],
pers=(0.0, 0.0),
means=meanloc,
stds=[self.std, self.std])
for x, y in zip(
np.random.normal(loc=meanloc[0],
scale=self.std,
size=self.npoints),
np.random.normal(loc=meanloc[1],
scale=self.std,
size=self.npoints)):
out.points.append(Pt(point=(x, y)))
return out
def trackSensors(self, sensors):
if sensors.bumpLeft or sensors.bumpRight:
# Set bump time (when it's ok to pay attention to the phi again)
self.bumpFreeTime = rospy.Time.now().to_sec() + 1.0
# Execute evasive action.
evade = Twist()
evade.linear.x = -self.maxSpeed
self.velPub.publish(evade)
# Register the obstacle with the phi function.
obstaclePos = (self.pos[0] + 0.01 * math.cos(self.pos[2]),
self.pos[1] + 0.01 * math.sin(self.pos[2]), 0.0)
# print self.name, "found: ", obstaclePos
# print self.name, "thinks he's at: ", self.pos
self.phi.addObstacle(obstaclePos)
# self.goal = self.pos
self.nextMoveTime = self.bumpFreeTime #rospy.Time.now().to_sec()
def trackOdom(self, req):
"""
This gets called as often as a new odom is issued. In the field we'd
be updating self.pos with position estimates from the perfesser, but
this is a model so we're cheating.
"""
if rospy.Time.now().to_sec() > self.bumpFreeTime:
self.bumpFreeTime = 0.0
sz = req.pose.pose.orientation.z
sw = req.pose.pose.orientation.w
th = math.copysign(2 * math.asin(sz), sz * sw)
self.pos = (req.pose.pose.position.x, req.pose.pose.position.y, th)
self.phi.updateOccupancy(self.pos)
self.posPub.publish(self.getPosEst())
now = rospy.Time.now().to_sec()
if self.nextMoveTime - now < self.movePeriod:
self.evaluateMove()
self.nextMoveTime = now + self.movePeriod
if rospy.Time.now().to_sec() > self.bumpFreeTime:
self.velPub.publish(self.stepThatWay())
def takeAdvice(self, req):
"""
Takes a piece of advice from its neighbor, rates it, and
returns the rating. The advice is a distribution of points to
which the other robot thinks this robot ought to go. We
compare that advice to our own intentions and return ratings
indicating the degree of agreement between the two: 1.0 is
complete agreement and 0.0 is complete disagreement. We never
return exactly 0.0 as a matter of policy.
"""
if not req.question:
return MRFQueryResponse(no_data=True)
if not self.deltaConfig.any():
return MRFQueryResponse(no_data=True)
# Allocate our intention into bins.
question = self.ptsToArray(req.question)
# Establish bins so that there is one point in each outer bin.
# (The Histogram class in perfesser is set up so the outer
# bins will be empty, which is why we're not using it here.
mins = [min(self.deltaConfig[:, 0]), min(self.deltaConfig[:, 1])]
maxs = [max(self.deltaConfig[:, 0]), max(self.deltaConfig[:, 1])]
bins = [ np.linspace(mn + 1.e-10, mx - 1.e-10, bn - 1) for \
mn, mx, bn in zip(mins, maxs, self.nbins) ]
hist = np.zeros(self.nbins)
# Create a distribution of points that includes neighbors'
# opinions, omitting the one asking the question.
phi = self.deltaConfig[:, (0, 1)]
messageProduct = [1.0] * phi.shape[0]
for nb in self.neighbors:
if (nb.name != req.asker) and (nb.lastReply.points):
messageProduct = [ m * p.point[0] for m,p in \
zip(messageProduct,
nb.lastReply.points) ]
phi = self.resample(phi, messageProduct)
# print "phi", phi
# Populate bins
for delta in phi:
hist[sum(delta[0] > bins[0])][sum(delta[1] > bins[1])] += 1.0
n = sum(sum(hist))
# Formulate a reply by binning the input data.
m = MRFQueryResponse(no_data=False, source_stamp=req.source_stamp)
for q in req.question:
m.reply.append(Pt(point=( \
hist[sum(q.point[0] > bins[0])][sum(q.point[1] > bins[1])]/n, )))
# print "entering service: ", self.name
# print self.name, "'s opinion: ", self.deltaConfig
# print "question: ",question
# print "hist: ", hist
# print "bins: ", bins
# print m
return m
def ptsToArray(self, pts):
return np.array([p.point for p in pts])
def arrayToPts(self, array):
return [Pt(point=a) for a in array]
def setAdvice(self, name, advice):
"""
Takes a name of a robot and an array of advice and stores it
in the neighbor array -- if appropriate.
"""
# print "setAdvice:", name, advice
for nb in self.neighbors:
if nb.name == name:
# print "storing............................"
nb.lastQuestion.points = self.arrayToPts(advice)
def evaluateMove(self):
"""
Formulates a distribution of points representing a belief
about the direction in which we should move, along with our
neighbors. The belief is based on our measurements of our own
location, as well as the location of the other robots, along
with some goals (i.e. spread out, explore, etc).
To be explicit, inputs are these:
- Location of self
- Location of neighbors
- Suggestions from neighbors about where self should move.
- Replies by neighbors to suggestions about where they
should move.
The first two form the input to the phi() function. The third
is the input to the psi() function, and the fourth is the
messageProduct referenced below.
Definition: "configuration space" is the n-dimensional space
where each point represents a particular configuration of all
the robots in it. So a one-dimensional room containing three
robots corresponds to a three-dimensional configuration
space. A two-dimensional room with five robots makes ten
dimensions in configuration space. Seems confusing, but it
makes some of the math easier.
"""
if self.busy or (self.bumpFreeTime > 0.0):
return
self.busy = True
# Marshal the inputs: our position and the positions of our
# neighbors. Park them all in an array where each row
# represents a position in configuration space and the
# collection of rows represent the distribution of those
# points.
currentConfig = np.array(self.ptsToArray(self.getPosEst().points))
# For 2d this should be two entries with X and Y
self.deltaNames = [self.name + "/dx", self.name + "/dy"]
for nb in self.neighbors:
if nb.location.points:
currentConfig = np.hstack(
[currentConfig,
self.ptsToArray(nb.location.points)])
self.deltaNames += [nb.name + "/dx", nb.name + "/dy"]
# Run through the points in the current configuration and
# generate predictions for each point in the distribution.
deltaPts = []
for cpoint in currentConfig:
deltaPts.append(self.phi.phi(cpoint))
self.deltaConfig = np.array(deltaPts)
# Break up the array, and store the opinions about each
# neighbor in the neighbor list. We do it this way in case
# the neighbor list has changed since the marshaling.
# print "names:", self.deltaNames
# print "shape:", self.deltaConfig.shape
# print "data: ", self.deltaConfig
machineNames = [ self.deltaNames[i].split("/")[0] \
for i in range(0, len(self.deltaNames), 2) ]
for nm, dc in zip(
machineNames,
np.hsplit(self.deltaConfig, self.deltaConfig.shape[1] / 2)):
# Record the advice given to that robot.
self.setAdvice(nm, dc)
# Notify each neighbor of our opinion (and record their
# responses).
messageProduct = [1.0] * self.npoints
for nb in self.neighbors:
# If we don't already have an opinion about where this
# robot should be then pass for the moment and we'll get
# it the next time around.
if not nb.lastQuestion.points:
break
m = MRFQueryRequest(source_stamp=rospy.Time.now(),
asker=self.name,
asked=nb.name,
question=nb.lastQuestion.points)
r = nb.advise(m)
if r.reply:
nb.lastReply.points = r.reply
nb.lastReply.source_stamp = r.source_stamp
messageProduct = [ lr.point[0] * mp \
for lr, mp in zip(r.reply,
messageProduct) ]
# Resample based on the neighbor's responses.
desiredConfig = self.resample(self.deltaConfig, messageProduct)
# The first set of coordinates in this is the desired place
# for us to move.
moves = desiredConfig[:, (0, 1)]
# print self.dispConf("now",currentConfig)
# print self.dispConf("own",self.deltaConfig)
# print self.dispConf("new",desiredConfig)
self.goThatWay(moves)
self.busy = False
def dispConf(self, name, configArray):
out = []
for row in configArray.transpose():
out.append(np.mean(row))
fmt = name + " " + self.name + ": (" + "%.3f," * len(out) + ")"
return fmt % tuple(out)
def goThatWay(self, moves):
"""
Takes an array of points to move to and selects a direction in
which to move in order to get us there.
"""
self.goal = (self.pos[0] + np.mean(moves[:, 0]),
self.pos[1] + np.mean(moves[:, 1]), 0.0)
def stepThatWay(self):
fX = self.goal[0] - self.pos[0]
fY = self.goal[1] - self.pos[1]
fM = math.hypot(fX, fY)
fD = math.atan2(fY, fX)
# Convert to robot frame
fDR = fD - self.pos[2]
if math.fabs(fDR) > math.pi:
fDR -= math.copysign(2 * math.pi, fDR)
if math.fabs(fDR) > math.pi / 10.0:
provX = 0.0
else:
provX = min(self.maxSpeed, self.kSpeed * fM)
provZ = math.fabs(self.kTurn * fDR)
provZ = min(self.maxTurn, provZ)
provZ = math.copysign(provZ, fDR)
twist = Twist()
twist.linear.x = provX
twist.angular.z = provZ
return twist
def weightedSample(self, pointProbs, numSamples):
"""
The following sampling algorithm is adapted from a sneaky
algorithm for coming up with a selection of N random numbers
between 0 and Q. The idea is that the probability that all N
of the numbers are less than some X is P=(X/Q)**N. Solve for
X in terms of P, and you get an equation that will pick the
largest number in that distribution for some random P. Do
that N times, and you get N samples.
The input is an indexed array of probabilities (use
enumerate(P) for the input, e.g. and the number of samples to
take from that array. The output is an iterator that will
provide the indices you can use to sample the original array
with, as in:
output = locArray[weightedSample(enumerate(locProbs), n)
This won't exactly work, since the enumerate has to be turned
into a sequence before it can be subscripted (as below) and
the same is true of the function output.
For more: stackoverflow.com/questions/2140787
"""
total = sum(prob for i, prob in pointProbs)
assert (not math.isnan(total))
j = 0
i, prob = pointProbs[j]
while numSamples:
x = total * (1 - np.random.random()**(1.0 / numSamples))
total -= x
while x > prob:
x -= prob
j += 1
i, prob = pointProbs[j]
prob -= x
yield i
numSamples -= 1
def resample(self, array, ratings):
"""
Resamples the distribution represented by the samples in
array, based on the list of ratings. There should be one
rating for each row of the array, and each rating should be
greater than zero and less than or equal to 1.0. The ratings
list need not be normed.
"""
#print "resample according to: ", ratings
newseq = self.weightedSample([(i, p) for i, p in enumerate(ratings)],
len(ratings))
#print "comes up with: ", [p for p in newseq]
wiggle = lambda x: self.std * (np.random.random() - 0.5) + x
out = []
for pt in array[[i for i in newseq], :]:
out.append([wiggle(pt[0]), wiggle(pt[1])])
return np.array(out)
class coverControl(object):
def __init__(self, name, neighborDist = 2.0, npoints = 70):
print "Starting." , name
self.name = name
self.npoints = npoints
self.neighborDist = neighborDist
# This is a list of the robots on our team, assumed to contain
# Neighbor() objects, as above.
self.neighbors = []
# This is a list of the robots within neighborDist of us.
self.neighborNames = []
# This is an array of points in configuration space indicating
# where we need to go next.
self.deltaConfig = np.array([[]])
self.twist = Twist()
self.maxSpeed = 0.5 # 0.25
self.maxTurn = 1.0 # .6
self.kSpeed = 15.0
self.kTurn = 13.0
# These are coefficients for the opinion formation. Tune and forget.
self.aw = 0.5
self.sp = 0.2 # Without this term, you get an exploding ring more often.
self.av = 0.6
self.se = 0.0
self.maxDist = 2.5
self.phi = Phi(self.name,
self.aw, self.sp, self.av, self.se,
self.maxDist)
self.movePeriod = 5.0
self.nextMoveTime = rospy.Time.now().to_sec() + self.movePeriod
# Estimated variance in the position measurements.
self.std = 0.01
# When rating a neighbor's opinion about our desired position,
# use this many bins. Shouldn't be a radically different
# order of magnitude from npoints.
self.nbins = (5,5)
rospy.sleep(1.0)
self.odomSub = rospy.Subscriber(self.name + "/odom", Odometry,
self.trackOdom,
queue_size = 1, buff_size = 250)
self.advice = rospy.Service(self.name + "/advice", MRFQuery,
self.takeAdvice)
self.busy = False
def addNeighbor(self, name):
self.neighbors.append(Neighbor(name))
def addNeighbors(self, names):
for name in names:
self.addNeighbor(name)
def recordNeighbors(self):
"""
At any time, the self.neighbors list should contain an entry for
all the robots in the experiment and the self.nbrNames list should
contain the names of the robots whose positions are close enough to
consider true neighbors.
"""
self.neighborNames = []
for nb in self.neighbors:
if self.dist(self.pos, nb.location.means) < self.neighborDist:
self.neighborNames.append(nb.name)
if not self.neighborNames:
# No neighbors -- should act on that. For now just
# return.
return
def trackBump(self, req):
self.phi.addObstacle(req.means)
def trackPos(self, req):
"""
This gets called as often as a new position estimate is issued.
"""
self.posBelief = req
self.pos = req.means
self.phi.updateOccupancy(self.pos)
now = rospy.Time.now().to_sec()
if self.nextMoveTime - now < self.movePeriod:
self.evaluateMove()
self.nextMoveTime = now + self.movePeriod
def takeAdvice(self, req):
"""
Takes a piece of advice from its neighbor, rates it, and
returns the rating. The advice is a distribution of points to
which the other robot thinks this robot ought to go. We
compare that advice to our own intentions and return ratings
indicating the degree of agreement between the two: 1.0 is
complete agreement and 0.0 is complete disagreement. We never
return exactly 0.0 as a matter of policy.
"""
if not req.question:
return MRFQueryResponse(no_data = True)
if not self.deltaConfig.any():
return MRFQueryResponse(no_data = True)
# Allocate our intention into bins.
question = self.ptsToArray(req.question)
# Establish bins so that there is one point in each outer bin.
# (The Histogram class in perfesser is set up so the outer
# bins will be empty, which is why we're not using it here.
mins = [ min(self.deltaConfig[:,0]), min(self.deltaConfig[:,1]) ]
maxs = [ max(self.deltaConfig[:,0]), max(self.deltaConfig[:,1]) ]
bins = [ np.linspace(mn + 1.e-10, mx - 1.e-10, bn) for \
mn,mx,bin in zip(mins, maxs, self.nbins) ]
hist = np.zeros(self.nbins)
# Create a distribution of points that includes neighbor's
# opinions except for the one asking the question.
phi = self.deltaConfig[:,(0,1)]
messageProduct = [ 1.0 ] * phi.shape[0]
for nb in self.neighbors:
if (nb.name != req.asker) and (nb.lastReply.points):
messageProduct = [ m * p.point[0] for m,p in \
zip(messageProduct,
nb.lastReply.points) ]
phi = self.resample(phi, messageProduct)
# Populate bins
for delta in phi:
hist[sum(delta[0] > bins[0])][sum(delta[1] > bins[1])] += 1.0
n = sum(sum(hist))
# Formulate a reply by binning the input data.
m = MRFQueryResponse(no_data = False,
source_stamp = req.source_stamp)
for q in req.question:
m.reply.append(Pt(point=( \
hist[sum(q.point[0] > bins[0])][sum(q.point[1] > bins[1])]/n, )))
# print "entering service: ", self.name
# print self.name, "'s opinion: ", self.deltaConfig
# print "question: ",question
# print "hist: ", hist
# print "bins: ", bins
# print m
return m
def ptsToArray(self, pts):
return np.array([p.point for p in pts])
def arrayToPts(self, array):
return [ Pt(point = a) for a in array ]
def setAdvice(self, name, advice):
"""
Takes a name of a robot and an array of advice and stores it
in the neighbor array -- if appropriate.
"""
for nb in self.neighbors:
if nb.name == name:
nb.lastQuestion.points = self.arrayToPts(advice)
def evaluateMove(self):
"""
Formulates a distribution of points representing a belief
about the direction in which we should move, along with our
neighbors. The belief is based on our measurements of our own
location, as well as the location of the other robots, along
with some goals (i.e. spread out, explore, etc).
To be explicit, inputs are these:
- Location of self
- Location of neighbors
- Suggestions from neighbors about where self should move.
- Replies by neighbors to suggestions about where they
should move.
The first two form the input to the phi() function. The third
is the input to the psi() function, and the fourth is the
messageProduct referenced below.
Definition: "configuration space" is the n-dimensional space
where each point represents a particular configuration of all
the robots in it. So a one-dimensional room containing three
robots corresponds to a three-dimensional configuration
space. A two-dimensional room with five robots makes ten
dimensions in configuration space. Seems confusing, but it
makes some of the math easier.
"""
# Marshal the inputs: our position and the positions of our
# neighbors. Park them all in an array where each row
# represents a position in configuration space and the
# collection of rows represent the distribution of those
# points.
currentConfig = np.array(self.ptsToArray(self.posBelief.points))[:,(0,1)]
# For 2d this should be two entries with X and Y
self.deltaNames = [ self.name + "/dx", self.name + "/dy" ]
# All the robots are in the neighbors list, but only the nearby ones
# are to be used.
for nb in self.neighbors:
if nb.location.points:
currentConfig = np.hstack([currentConfig,
self.ptsToArray(nb.location.points)])
self.deltaNames += [ nb.name + "/dx", nb.name + "/dy" ]
# Run through the points in the current configuration and
# generate predictions for each point in the distribution.
deltaPts = []
for cpoint in currentConfig:
deltaPts.append(self.phi.phi(cpoint))
self.deltaConfig = np.array(deltaPts)
# Break up the array, and store the opinions about each
# neighbor in the neighbor list. We do it this way in case
# the neighbor list has changed since the marshaling.
for nm,dc in zip(self.deltaNames,np.hsplit(self.deltaConfig,
self.deltaConfig.shape[1])):
self.setAdvice(nm, dc)
# Notify each neighbor of our opinion (and record their
# responses).
messageProduct = [ 1.0 ] * self.npoints
for nb in self.neighbors:
# If we don't already have an opinion about where this
# robot should be then pass for the moment and we'll get
# it the next time around.
if not nb.lastQuestion.points:
break
m = MRFQueryRequest(source_stamp = rospy.Time.now(),
asker = self.name,
asked = nb.name,
question = nb.lastQuestion.points)
r = nb.advise(m)
if r.reply:
nb.lastReply.points = r.reply
nb.lastReply.source_stamp = r.source_stamp
messageProduct = [ lr.point[0] * mp \
for lr, mp in zip(r.reply,
messageProduct) ]
# Resample based on the neighbor's responses.
desiredConfig = self.resample(self.deltaConfig, messageProduct)
# The first set of coordinates in this is the desired place
# for us to move.
moves = desiredConfig[:,(0,1)]
self.goThatWay(moves)
self.busy = False
def goThatWay(self, moves):
"""
Takes an array of points to move to and selects a direction in
which to move in order to get us there.
"""
self.goal = (self.pos[0] + np.mean(moves[:, 0]),
self.pos[1] + np.mean(moves[:, 1]), 0.0)
####### SET NEW GOAL HERE #######
def weightedSample(self, pointProbs, numSamples):
"""
The following sampling algorithm is adapted from a sneaky
algorithm for coming up with a selection of N random numbers
between 0 and Q. The idea is that the probability that all N
of the numbers are less than some X is P=(X/Q)**N. Solve for
X in terms of P, and you get an equation that will pick the
largest number in that distribution for some random P. Do
that N times, and you get N samples.
The input is an indexed array of probabilities (use
enumerate(P) for the input, e.g. and the number of samples to
take from that array. The output is an iterator that will
provide the indices you can use to sample the original array
with, as in:
output = locArray[weightedSample(enumerate(locProbs), n)
This won't exactly work, since the enumerate has to be turned
into a sequence before it can be subscripted (as below) and
the same is true of the function output.
For more: stackoverflow.com/questions/2140787
"""
total = sum(prob for i, prob in pointProbs)
assert(not math.isnan(total))
j = 0
i, prob = pointProbs[j]
while numSamples:
x = total * (1 - np.random.random() ** (1.0/numSamples))
total -= x
while x > prob:
x -= prob
j += 1
i, prob = pointProbs[j]
prob -= x
yield i
numSamples -= 1
def resample(self, array, ratings):
"""
Resamples the distribution represented by the samples in
array, based on the list of ratings. There should be one
rating for each row of the array, and each rating should be
greater than zero and less than or equal to 1.0. The ratings
list need not be normed.
"""
#print "resample according to: ", ratings
newseq = self.weightedSample([(i,p) for i,p in enumerate(ratings)],
len(ratings))
#print "comes up with: ", [p for p in newseq]
wiggle = lambda x: self.std * (np.random.random() - 0.5) + x
out = []
for pt in array[ [ i for i in newseq ], :]:
out.append([ wiggle(pt[0]), wiggle(pt[1]) ])
return np.array(out)
def __init__(self, name, initpos=(0.0,0.0,0.0), npoints = 10):
print "Starting." , name
self.name = name
self.pos = initpos
self.npoints = npoints
# This is a list of neighboring robots, assumed to contain
# Neighbor() objects, as above.
self.neighbors = []
self.velPub = rospy.Publisher(self.name + "/cmd_vel", Twist)
self.posPub = rospy.Publisher(self.name + "/position", Belief)
# This is an array of points in configuration space indicating
# where we need to go next.
self.deltaConfig = np.array([[]])
self.twist = Twist()
self.maxSpeed = 0.5 # 0.25
self.maxTurn = 1.0 # .6
self.kSpeed = 15.0
self.kTurn = 13.0
# These are coefficients for the opinion formation.
self.aw = 13.
self.sp = 0.1 # Without this term, you get an exploding ring more often.
self.av = 3.
self.se = 250.0
self.maxDist = 2.5
self.phi = Phi(self.name,
self.aw, self.sp, self.av, self.se,
self.maxDist)
self.movePeriod = 5.0
self.nextMoveTime = rospy.Time.now().to_sec() + self.movePeriod
# When this is in the future, we are still dealing with a bump.
self.bumpFreeTime = 0.0
# Estimated variance in the position measurements.
self.std = 0.01
# When rating a neighbor's opinion about our desired position,
# use this many bins. Shouldn't be a radically different
# order of magnitude from npoints.
self.nbins = (5,5)
rospy.sleep(1.0)
self.sensorSub = rospy.Subscriber(self.name + "/sensors", SensorPacket,
self.trackSensors)
self.odomSub = rospy.Subscriber(self.name + "/odom", Odometry,
self.trackOdom,
queue_size = 1, buff_size = 250)
self.advice = rospy.Service(self.name + "/advice", MRFQuery,
self.takeAdvice)
self.roomLoc = rospy.Subscriber("/room/locations", Room2D,
self.recordNeighbors)
self.busy = False
class modelControl(object):
def __init__(self, name, initpos=(0.0,0.0,0.0), npoints = 10):
print "Starting." , name
self.name = name
self.pos = initpos
self.npoints = npoints
# This is a list of neighboring robots, assumed to contain
# Neighbor() objects, as above.
self.neighbors = []
self.velPub = rospy.Publisher(self.name + "/cmd_vel", Twist)
self.posPub = rospy.Publisher(self.name + "/position", Belief)
# This is an array of points in configuration space indicating
# where we need to go next.
self.deltaConfig = np.array([[]])
self.twist = Twist()
self.maxSpeed = 0.5 # 0.25
self.maxTurn = 1.0 # .6
self.kSpeed = 15.0
self.kTurn = 13.0
# These are coefficients for the opinion formation.
self.aw = 13.
self.sp = 0.1 # Without this term, you get an exploding ring more often.
self.av = 3.
self.se = 250.0
self.maxDist = 2.5
self.phi = Phi(self.name,
self.aw, self.sp, self.av, self.se,
self.maxDist)
self.movePeriod = 5.0
self.nextMoveTime = rospy.Time.now().to_sec() + self.movePeriod
# When this is in the future, we are still dealing with a bump.
self.bumpFreeTime = 0.0
# Estimated variance in the position measurements.
self.std = 0.01
# When rating a neighbor's opinion about our desired position,
# use this many bins. Shouldn't be a radically different
# order of magnitude from npoints.
self.nbins = (5,5)
rospy.sleep(1.0)
self.sensorSub = rospy.Subscriber(self.name + "/sensors", SensorPacket,
self.trackSensors)
self.odomSub = rospy.Subscriber(self.name + "/odom", Odometry,
self.trackOdom,
queue_size = 1, buff_size = 250)
self.advice = rospy.Service(self.name + "/advice", MRFQuery,
self.takeAdvice)
self.roomLoc = rospy.Subscriber("/room/locations", Room2D,
self.recordNeighbors)
self.busy = False
def recordNeighbors(self, req):
"""
At any time, the self.neighbors list should contain a Neighbor
object for each robot whose position we can measure directly.
For the purpose of simulation, we are receiving a list of
robots and positions via the /room/locations topic and
accumulating. Use this as a callback to the /room/locations
topic. The result of calling this is an updated
self.neighbors array.
"""
distances = []
nabe = False
for nb in req.robots:
if nb.sourceName == self.name:
nabe = nb
break
if not nabe:
# No neighbors -- should act on that. For now just
# return.
return
# Now we have a list of robots and locations that are neighbors
# Check them against the existing list and edit accordingly.
nameList = [ nb.name for nb in self.neighbors ]
for name in nabe.names:
if not (name in nameList):
self.neighbors.append(Neighbor(name = name))
for nb in self.neighbors:
if not (nb.name in nabe.names):
self.neighbors.remove(nb)
def getPosEst(self):
meanloc = ( np.random.normal(loc = self.pos[0], scale = self.std/2.,
size = 1)[0],
np.random.normal(loc = self.pos[1], scale = self.std/2.,
size = 1)[0])
out = Belief(source_stamp = rospy.Time.now(),
source_data = "estimate",
data_type = "location",
data_sub_type = self.name,
sender = self.name,
dim_names = [ self.name + "/xpos",
self.name + "/ypos" ],
pers = (0.0,0.0),
means = meanloc,
stds = [ self.std, self.std ])
for x,y in zip(np.random.normal(loc = meanloc[0], scale = self.std,
size = self.npoints),
np.random.normal(loc = meanloc[1], scale = self.std,
size = self.npoints)):
out.points.append(Pt(point = (x,y)))
return out
def trackSensors(self, sensors):
if sensors.bumpLeft or sensors.bumpRight:
# Set bump time (when it's ok to pay attention to the phi again)
self.bumpFreeTime = rospy.Time.now().to_sec() + 1.0
# Execute evasive action.
evade = Twist()
evade.linear.x = -self.maxSpeed
self.velPub.publish(evade)
# Register the obstacle with the phi function.
obstaclePos = ( self.pos[0] + 0.01 * math.cos(self.pos[2]),
self.pos[1] + 0.01 * math.sin(self.pos[2]),
0.0 )
# print self.name, "found: ", obstaclePos
# print self.name, "thinks he's at: ", self.pos
self.phi.addObstacle(obstaclePos)
# self.goal = self.pos
self.nextMoveTime = self.bumpFreeTime #rospy.Time.now().to_sec()
def trackOdom(self, req):
"""
This gets called as often as a new odom is issued. In the field we'd
be updating self.pos with position estimates from the perfesser, but
this is a model so we're cheating.
"""
if rospy.Time.now().to_sec() > self.bumpFreeTime:
self.bumpFreeTime = 0.0
sz = req.pose.pose.orientation.z
sw = req.pose.pose.orientation.w
th = math.copysign(2 * math.asin(sz), sz*sw)
self.pos = (req.pose.pose.position.x,
req.pose.pose.position.y,
th)
self.phi.updateOccupancy(self.pos)
self.posPub.publish(self.getPosEst())
now = rospy.Time.now().to_sec()
if self.nextMoveTime - now < self.movePeriod:
self.evaluateMove()
self.nextMoveTime = now + self.movePeriod
if rospy.Time.now().to_sec() > self.bumpFreeTime:
self.velPub.publish(self.stepThatWay())
def takeAdvice(self, req):
"""
Takes a piece of advice from its neighbor, rates it, and
returns the rating. The advice is a distribution of points to
which the other robot thinks this robot ought to go. We
compare that advice to our own intentions and return ratings
indicating the degree of agreement between the two: 1.0 is
complete agreement and 0.0 is complete disagreement. We never
return exactly 0.0 as a matter of policy.
"""
if not req.question:
return MRFQueryResponse(no_data = True)
if not self.deltaConfig.any():
return MRFQueryResponse(no_data = True)
# Allocate our intention into bins.
question = self.ptsToArray(req.question)
# Establish bins so that there is one point in each outer bin.
# (The Histogram class in perfesser is set up so the outer
# bins will be empty, which is why we're not using it here.
mins = [ min(self.deltaConfig[:,0]), min(self.deltaConfig[:,1]) ]
maxs = [ max(self.deltaConfig[:,0]), max(self.deltaConfig[:,1]) ]
bins = [ np.linspace(mn + 1.e-10, mx - 1.e-10, bn - 1) for \
mn, mx, bn in zip(mins, maxs, self.nbins) ]
hist = np.zeros(self.nbins)
# Create a distribution of points that includes neighbors'
# opinions, omitting the one asking the question.
phi = self.deltaConfig[:,(0,1)]
messageProduct = [ 1.0 ] * phi.shape[0]
for nb in self.neighbors:
if (nb.name != req.asker) and (nb.lastReply.points):
messageProduct = [ m * p.point[0] for m,p in \
zip(messageProduct,
nb.lastReply.points) ]
phi = self.resample(phi, messageProduct)
# print "phi", phi
# Populate bins
for delta in phi:
hist[sum(delta[0] > bins[0])][sum(delta[1] > bins[1])] += 1.0
n = sum(sum(hist))
# Formulate a reply by binning the input data.
m = MRFQueryResponse(no_data = False,
source_stamp = req.source_stamp)
for q in req.question:
m.reply.append(Pt(point=( \
hist[sum(q.point[0] > bins[0])][sum(q.point[1] > bins[1])]/n, )))
# print "entering service: ", self.name
# print self.name, "'s opinion: ", self.deltaConfig
# print "question: ",question
# print "hist: ", hist
# print "bins: ", bins
# print m
return m
def ptsToArray(self, pts):
return np.array([p.point for p in pts])
def arrayToPts(self, array):
return [ Pt(point = a) for a in array ]
def setAdvice(self, name, advice):
"""
Takes a name of a robot and an array of advice and stores it
in the neighbor array -- if appropriate.
"""
# print "setAdvice:", name, advice
for nb in self.neighbors:
if nb.name == name:
# print "storing............................"
nb.lastQuestion.points = self.arrayToPts(advice)
def evaluateMove(self):
"""
Formulates a distribution of points representing a belief
about the direction in which we should move, along with our
neighbors. The belief is based on our measurements of our own
location, as well as the location of the other robots, along
with some goals (i.e. spread out, explore, etc).
To be explicit, inputs are these:
- Location of self
- Location of neighbors
- Suggestions from neighbors about where self should move.
- Replies by neighbors to suggestions about where they
should move.
The first two form the input to the phi() function. The third
is the input to the psi() function, and the fourth is the
messageProduct referenced below.
Definition: "configuration space" is the n-dimensional space
where each point represents a particular configuration of all
the robots in it. So a one-dimensional room containing three
robots corresponds to a three-dimensional configuration
space. A two-dimensional room with five robots makes ten
dimensions in configuration space. Seems confusing, but it
makes some of the math easier.
"""
if self.busy or (self.bumpFreeTime > 0.0):
return
self.busy = True
# Marshal the inputs: our position and the positions of our
# neighbors. Park them all in an array where each row
# represents a position in configuration space and the
# collection of rows represent the distribution of those
# points.
currentConfig = np.array(self.ptsToArray(self.getPosEst().points))
# For 2d this should be two entries with X and Y
self.deltaNames = [ self.name + "/dx", self.name + "/dy" ]
for nb in self.neighbors:
if nb.location.points:
currentConfig = np.hstack([currentConfig,
self.ptsToArray(nb.location.points)])
self.deltaNames += [ nb.name + "/dx", nb.name + "/dy" ]
# Run through the points in the current configuration and
# generate predictions for each point in the distribution.
deltaPts = []
for cpoint in currentConfig:
deltaPts.append(self.phi.phi(cpoint))
self.deltaConfig = np.array(deltaPts)
# Break up the array, and store the opinions about each
# neighbor in the neighbor list. We do it this way in case
# the neighbor list has changed since the marshaling.
# print "names:", self.deltaNames
# print "shape:", self.deltaConfig.shape
# print "data: ", self.deltaConfig
machineNames = [ self.deltaNames[i].split("/")[0] \
for i in range(0, len(self.deltaNames), 2) ]
for nm,dc in zip(machineNames,
np.hsplit(self.deltaConfig,
self.deltaConfig.shape[1]/2)):
# Record the advice given to that robot.
self.setAdvice(nm, dc)
# Notify each neighbor of our opinion (and record their
# responses).
messageProduct = [ 1.0 ] * self.npoints
for nb in self.neighbors:
# If we don't already have an opinion about where this
# robot should be then pass for the moment and we'll get
# it the next time around.
if not nb.lastQuestion.points:
break
m = MRFQueryRequest(source_stamp = rospy.Time.now(),
asker = self.name,
asked = nb.name,
question = nb.lastQuestion.points)
r = nb.advise(m)
if r.reply:
nb.lastReply.points = r.reply
nb.lastReply.source_stamp = r.source_stamp
messageProduct = [ lr.point[0] * mp \
for lr, mp in zip(r.reply,
messageProduct) ]
# Resample based on the neighbor's responses.
desiredConfig = self.resample(self.deltaConfig, messageProduct)
# The first set of coordinates in this is the desired place
# for us to move.
moves = desiredConfig[:,(0,1)]
# print self.dispConf("now",currentConfig)
# print self.dispConf("own",self.deltaConfig)
# print self.dispConf("new",desiredConfig)
self.goThatWay(moves)
self.busy = False
def dispConf(self, name, configArray):
out = []
for row in configArray.transpose():
out.append(np.mean(row))
fmt = name + " " + self.name + ": (" + "%.3f," * len(out) + ")"
return fmt % tuple(out)
def goThatWay(self, moves):
"""
Takes an array of points to move to and selects a direction in
which to move in order to get us there.
"""
self.goal = (self.pos[0] + np.mean(moves[:, 0]),
self.pos[1] + np.mean(moves[:, 1]), 0.0)
def stepThatWay(self):
fX = self.goal[0] - self.pos[0]
fY = self.goal[1] - self.pos[1]
fM = math.hypot(fX, fY)
fD = math.atan2(fY, fX)
# Convert to robot frame
fDR = fD - self.pos[2]
if math.fabs(fDR) > math.pi:
fDR -= math.copysign(2 * math.pi, fDR)
if math.fabs(fDR) > math.pi/10.0:
provX = 0.0
else:
provX = min(self.maxSpeed, self.kSpeed * fM)
provZ = math.fabs(self.kTurn * fDR)
provZ = min(self.maxTurn, provZ)
provZ = math.copysign(provZ, fDR)
twist = Twist()
twist.linear.x = provX
twist.angular.z = provZ
return twist
def weightedSample(self, pointProbs, numSamples):
"""
The following sampling algorithm is adapted from a sneaky
algorithm for coming up with a selection of N random numbers
between 0 and Q. The idea is that the probability that all N
of the numbers are less than some X is P=(X/Q)**N. Solve for
X in terms of P, and you get an equation that will pick the
largest number in that distribution for some random P. Do
that N times, and you get N samples.
The input is an indexed array of probabilities (use
enumerate(P) for the input, e.g. and the number of samples to
take from that array. The output is an iterator that will
provide the indices you can use to sample the original array
with, as in:
output = locArray[weightedSample(enumerate(locProbs), n)
This won't exactly work, since the enumerate has to be turned
into a sequence before it can be subscripted (as below) and
the same is true of the function output.
For more: stackoverflow.com/questions/2140787
"""
total = sum(prob for i, prob in pointProbs)
assert(not math.isnan(total))
j = 0
i, prob = pointProbs[j]
while numSamples:
x = total * (1 - np.random.random() ** (1.0/numSamples))
total -= x
while x > prob:
x -= prob
j += 1
i, prob = pointProbs[j]
prob -= x
yield i
numSamples -= 1
def resample(self, array, ratings):
"""
Resamples the distribution represented by the samples in
array, based on the list of ratings. There should be one
rating for each row of the array, and each rating should be
greater than zero and less than or equal to 1.0. The ratings
list need not be normed.
"""
#print "resample according to: ", ratings
newseq = self.weightedSample([(i,p) for i,p in enumerate(ratings)],
len(ratings))
#print "comes up with: ", [p for p in newseq]
wiggle = lambda x: self.std * (np.random.random() - 0.5) + x
out = []
for pt in array[ [ i for i in newseq ], :]:
out.append([ wiggle(pt[0]), wiggle(pt[1]) ])
return np.array(out)
