Plan Refinement

BayRoB can be used to refine robot plans by searching for a path from the current state to a desired, user-defined goal state. The result is a sequence of belief states representing the prospected outcomes of parameterized action steps that, when executed in order, will most likely lead to the desired goal state. The class bayrob.core.base.Search encapsulates the information required to perform the search, i.e. the initial (belief) state of the agent and a goal specification, each augmented by a certain tolerance, and the direction, the search is supposed to be performed in. The search specification is then passed to the bayrob.core.base.BayRoB, which triggers the search algorithm.

import os
from jpt.base.intervals import ContinuousSet
from bayrob.core.base import BayRoB, Search
from bayrob.utils import locs
preset = {
     "init": {
         'x_in': 62,
         'y_in': 74,
         'xdir_in': .3,
         'ydir_in': .9,
     },
     "init_tolerances": {
         'x_in': .1,
         'y_in': .1,
         'xdir_in': .01,
         'ydir_in': .01,
     },
     "goal": {
         'detected(milk)': {True},
     },
     "goal_tolerances": {},
     "bwd": True
 }

bayrob = BayRoB()
bayrob.adddatapath([os.path.join(locs.examples, 'demo', d) for d in os.listdir(os.path.join(locs.examples, 'demo'))])
allvars = bayrob.models['move'].variables + \
                  bayrob.models['turn'].variables + \
                  bayrob.models['perception'].variables
allvars_ = {v.name: v for v in allvars}

asr = Search()
asr.bwd = preset['bwd']
asr.init = {allvars_[k]: v for k, v in preset['init'].items()}
asr.init_tolerances = {allvars_[k]: v for k, v in preset['init_tolerances'].items()}
asr.goal = {allvars_[k]: v for k, v in preset['goal'].items()}
asr.goal_tolerances = {allvars_[k]: v for k, v in preset['goal_tolerances'].items()}

bayrob.query = asr
bayrob.search_astar()
seq = bayrob.result.result
print(seq)

The result of a BayRoB search is a sequence of belief states (bayrob.core.astar_jpt.State) represented by instances of the probability distribution classes jpt.distributions.univariate.numeric.Numeric, jpt.distributions.univariate.multinomial.Bool, and jpt.distributions.univariate.multinomial.Multinomial, for multiple variables (jpt.variables.Variable).