Source code for concepts.dm.crow.planners.regression_planning_impl.crow_regression_planner_dfs_v2

#! /usr/bin/env python3
# -*- coding: utf-8 -*-
# File   : crow_regression_planner_dfs_v2.py
# Author : Jiayuan Mao
# Email  : maojiayuan@gmail.com
# Date   : 03/21/2024
#
# This file is part of Project Concepts.
# Distributed under terms of the MIT license.

import jacinle

from typing import Optional, Union, Sequence, Tuple, List, Dict, NamedTuple
from concepts.dsl.constraint import ConstraintSatisfactionProblem
from concepts.dsl.dsl_types import Variable
from concepts.dsl.expression import ValueOutputExpression
from concepts.dsl.tensor_state import StateObjectReference

from concepts.dm.crow.crow_domain import CrowState
from concepts.dm.crow.controller import CrowControllerApplier, CrowControllerApplicationExpression
from concepts.dm.crow.behavior import CrowAchieveExpression, CrowBindExpression, CrowRuntimeAssignmentExpression, CrowAssertExpression
from concepts.dm.crow.behavior import CrowBehaviorOrderingSuite, CrowBehaviorApplicationExpression
from concepts.dm.crow.behavior_utils import match_applicable_behaviors, ApplicableBehaviorItem, execute_object_bind, format_behavior_program

from concepts.dm.crow.planners.regression_planning import CrowPlanningResult, CrowRegressionPlanner, ScopedCrowExpression
from concepts.dm.crow.planners.regression_utils import canonize_bounded_variables
from concepts.dm.crow.planners.regression_planning_impl.crow_regression_planner_dfs_v1_utils import execute_behavior_effect_batch, unique_results


class _DFSNode(NamedTuple):
    program: CrowBehaviorOrderingSuite
    state: CrowState
    csp: ConstraintSatisfactionProblem
    scopes: Dict[int, dict]
    latest_scope: int
    left_actions: Tuple[CrowControllerApplier, ...]
    right_statements: Optional[List[ScopedCrowExpression]] = None
    allow_promotable: bool = True
    depth: int = 0


[docs] class CrowRegressionPlannerDFSv2(CrowRegressionPlanner): def _post_init(self, max_search_depth: int = 25): self.max_search_depth = max_search_depth
[docs] def main_entry(self, program: CrowBehaviorOrderingSuite, minimize: Optional[ValueOutputExpression] = None) -> List[Tuple[CrowControllerApplier, ...]]: state = _DFSNode(program, self.state, ConstraintSatisfactionProblem(), {0: {}}, 0, tuple(), list(), allow_promotable=True, depth=0) results = self.dfs(state) return [x.controller_actions for x in results]
[docs] @jacinle.log_function(verbose=False) def dfs(self, state: _DFSNode) -> Sequence[CrowPlanningResult]: if state.depth >= self.max_search_depth: print('Max depth reached.') # import ipdb; ipdb.set_trace() return [] # left_actions_str = ' '.join([str(x) for x in state.left_actions]) if len(state.left_actions) > 0 else '<empty left>' # right_statements_str = ' '.join([f'{str(x)}@{i}' for x, i in state.right_statements]) if len(state.right_statements) > 0 else '<empty right>' # jacinle.log_function.print( # 'LEFT: ' + jacinle.colored(left_actions_str, 'green'), # 'RIGHT: ' + jacinle.colored(right_statements_str, 'yellow'), # 'SCOPE: ' + str(state.scopes), # sep='\n' # ) # jacinle.log_function.print('Program:', state.program) # jacinle.log_function.print('Scopes:', state.scopes) # jacinle.log_function.print('Depth', state.depth) # def _is_sequential_program(program: CrowActionOrderingSuite) -> bool: # if program.order is not CrowActionOrderingSuite.ORDER.SEQUENTIAL: # return False # for stmt in program.statements: # if isinstance(stmt, CrowActionOrderingSuite): # if not _is_sequential_program(stmt): # return False # return True # # if not _is_sequential_program(state.program): # print('Non-sequential program.') # import ipdb; ipdb.set_trace() # pass new_results = list() for left, stmt, scope_id in state.program.pop_right_statement(): new_results.extend(self._dfs_expand(state, left, stmt, scope_id)) # jacinle.log_function.print('New results:', [x.actions for x in new_results]) return unique_results(new_results)
def _dfs_expand(self, state: _DFSNode, left: CrowBehaviorOrderingSuite, stmt: Union[CrowAchieveExpression, CrowControllerApplicationExpression], scope_id: int) -> Sequence[CrowPlanningResult]: """Expand a statement. Args: state: the current search state. left: the left part of the program. stmt: the statement to be expanded. scope_id: the current scope id. """ new_results = list() if isinstance(stmt, CrowAchieveExpression): jacinle.log_function.print('Expanding:', stmt, 'with left:', format_behavior_program(left, state.scopes) if left is not None else None, 'and scope_id:', scope_id) # jacinle.log_function.print('Null goal expansion', stmt.goal) # Branch 1: the goal is directly satisfied at the after expanding the left part (null production). if left is None: last_results = [CrowPlanningResult(state.state, state.csp, state.left_actions, state.scopes)] else: last_results = self.dfs(_DFSNode( left, state.state, state.csp, state.scopes, state.latest_scope, state.left_actions, [ScopedCrowExpression(CrowAchieveExpression(stmt.goal, once=stmt.once), scope_id)] + state.right_statements, allow_promotable=state.allow_promotable, depth=state.depth + 1 )) for last_result in last_results: rv = self.executor.execute(stmt.goal, state=last_result.state, bounded_variables=canonize_bounded_variables(last_result.scopes, scope_id)) # jacinle.log_function.print('Last result:', last_result.actions, 'Null Eval =', bool(rv.item())) if bool(rv.item()): new_results.append(last_result) if len(new_results) > 0: return new_results # Branch 2: the goal is not directly satisfied. We need to refine the goal. all_matching = match_applicable_behaviors(self.executor.domain, self.state, stmt.goal, state.scopes[scope_id]) all_matching = list(all_matching) # matchings_str = [(x.action.name, x.bounded_variables) for x in all_matching] # jacinle.log_function.print(jacinle.tabulate(matchings_str, headers=['Action', 'Bounded Variables'])) for action_matching in all_matching: # jacinle.log_function.print('Trying action:', action_matching.action, 'with bounded variables:', action_matching.bounded_variables) # if action_matching.action.name == 'r_clear_from_holding': # import ipdb; ipdb.set_trace() new_results.extend(self._dfs_expand_action(state, last_results, left, action_matching, scope_id)) elif isinstance(stmt, CrowBehaviorApplicationExpression): if left is None: last_results = [CrowPlanningResult(state.state, state.csp, state.left_actions, state.scopes)] else: last_results = self.dfs(_DFSNode( left, state.state, state.csp, state.scopes, state.latest_scope, state.left_actions, [ScopedCrowExpression(stmt, scope_id)] + state.right_statements, allow_promotable=state.allow_promotable, depth=state.depth )) for last_result in last_results: argument_values = [self.executor.execute(x, state=last_result.state, bounded_variables=canonize_bounded_variables(last_result.scopes, scope_id)) for x in stmt.arguments] action_matching = ApplicableBehaviorItem(stmt.behavior, {k.name: v for k, v in zip(stmt.behavior.arguments, argument_values)}) new_results.extend(self._dfs_expand_action(state, [last_result], left, action_matching, scope_id)) elif isinstance(stmt, (CrowBindExpression, CrowRuntimeAssignmentExpression, CrowAssertExpression, CrowControllerApplicationExpression)): # These statements are "primitive" or "atomic". Therefore, all we need to do is to expand the left branch and then apply the primitive. if left is None: last_results = [CrowPlanningResult(state.state, state.csp, state.left_actions, state.scopes)] else: last_results = self.dfs(_DFSNode( left, state.state, state.csp, state.scopes, state.latest_scope, state.left_actions, [ScopedCrowExpression(stmt, scope_id)] + state.right_statements, allow_promotable=state.allow_promotable, depth=state.depth )) for last_result in last_results: new_results.extend(self._dfs_expand_primitive(last_result, stmt, scope_id)) else: raise ValueError(f'Unknown statement type: {stmt}') return new_results def _dfs_expand_action(self, state: _DFSNode, last_results: Sequence[CrowPlanningResult], left: Optional[CrowBehaviorOrderingSuite], action_matching: ApplicableBehaviorItem, scope_id: int) -> Sequence[CrowPlanningResult]: new_results = list() bounded_variables = action_matching.bounded_variables for var, value in bounded_variables.items(): if isinstance(value, Variable): bounded_variables[var] = value.clone_with_scope(scope_id) # Create a new scope for this subgoal refinement. new_scope_id = state.latest_scope + 1 program = action_matching.behavior.assign_body_program_scope(new_scope_id) preamble, promotable, rest = program.split_preamble_and_promotable() new_scopes = state.scopes.copy() new_scopes[new_scope_id] = bounded_variables.copy() if preamble is None: last_results = [CrowPlanningResult(state.state, state.csp, state.left_actions, new_scopes)] else: last_results = self.dfs(_DFSNode( CrowBehaviorOrderingSuite.make_sequential(preamble, variable_scope_identifier=new_scope_id), state.state, state.csp, new_scopes, new_scope_id, state.left_actions, [], allow_promotable=False, depth=state.depth + 1 )) if len(last_results) == 0: return [] new_promotable_results = list() if promotable is None: if left is None: new_promotable_results = last_results else: for last_result in last_results: new_promotable_results.extend(self.dfs(_DFSNode( left, last_result.state, last_result.csp, last_result.scopes, new_scope_id, last_result.controller_actions, [ScopedCrowExpression(rest, scope_id)] + state.right_statements, allow_promotable=True, depth=state.depth + 1 ))) else: if left is None: program = CrowBehaviorOrderingSuite.make_sequential(promotable, variable_scope_identifier=new_scope_id) else: program = CrowBehaviorOrderingSuite.make_unordered(left, CrowBehaviorOrderingSuite.make_sequential(promotable, variable_scope_identifier=new_scope_id)) jacinle.log_function.print('Making unordered program!') jacinle.log_function.print(format_behavior_program(program, last_results[0].scopes)) # import ipdb; ipdb.set_trace() # pass right_program = [ScopedCrowExpression(rest, scope_id), ScopedCrowExpression(action_matching.behavior, scope_id)] + state.right_statements for last_result in last_results: results = self.dfs(_DFSNode( program, last_result.state, last_result.csp, last_result.scopes, new_scope_id, last_result.controller_actions, right_program, allow_promotable=True, depth=state.depth + 1 )) new_promotable_results.extend(results) if len(new_promotable_results) == 0: return [] right_program = [ScopedCrowExpression(action_matching.behavior, scope_id)] + state.right_statements for last_result in new_promotable_results: results = self.dfs(_DFSNode( CrowBehaviorOrderingSuite.make_sequential(rest, variable_scope_identifier=new_scope_id), last_result.state, last_result.csp, last_result.scopes, new_scope_id, last_result.controller_actions, right_program, allow_promotable=True, depth=state.depth + 1 )) new_results.extend(execute_behavior_effect_batch(self.executor, results, action_matching.behavior, new_scope_id)) return new_results def _dfs_expand_primitive(self, last_result: CrowPlanningResult, stmt: Union[CrowBindExpression, CrowRuntimeAssignmentExpression, CrowAssertExpression, CrowControllerApplicationExpression], scope_id: int) -> Sequence[CrowPlanningResult]: """Apply a primitive statement on top of a particular planning result of the left branch.""" # jacinle.log_function.print('Expanding primitive:', stmt, 'with scope_id:', scope_id) if isinstance(stmt, CrowControllerApplicationExpression): argument_values = [self.executor.execute(x, state=last_result.state, bounded_variables=canonize_bounded_variables(last_result.scopes, scope_id)) for x in stmt.arguments] for i, argv in enumerate(argument_values): if isinstance(argv, StateObjectReference): argument_values[i] = argv.name return [CrowPlanningResult(last_result.state, last_result.csp, last_result.controller_actions + (CrowControllerApplier(stmt.controller, argument_values),), last_result.scopes)] elif isinstance(stmt, CrowBindExpression): if stmt.is_object_bind: new_results = list() for new_scope in execute_object_bind(self.executor, stmt, last_result.state, canonize_bounded_variables(last_result.scopes, scope_id)): new_scopes = last_result.scopes.copy() new_scopes[scope_id] = new_scope new_results.append(CrowPlanningResult(last_result.state, last_result.csp, last_result.controller_actions, new_scopes)) # jacinle.log_function.print(f'Object binding results for {stmt} under scope {canonize_bounded_variables(last_result.scopes, scope_id)}:', len(new_results), 'states.') return new_results else: raise NotImplementedError() elif isinstance(stmt, CrowRuntimeAssignmentExpression): rv = self.executor.execute(stmt.value, state=last_result.state, bounded_variables=canonize_bounded_variables(last_result.scopes, scope_id)) new_scopes = last_result.scopes.copy() new_scopes[scope_id] = last_result.scopes[scope_id].copy() new_scopes[scope_id][stmt.variable.name] = rv.item() return [CrowPlanningResult(last_result.state, last_result.csp, last_result.controller_actions, new_scopes)] elif isinstance(stmt, CrowAssertExpression): # TODO(Jiayuan Mao @ 2024/03/19): implement the CSP tracking. rv = self.executor.execute(stmt.bool_expr, state=last_result.state, bounded_variables=canonize_bounded_variables(last_result.scopes, scope_id)) # jacinle.log_function.print(f'Assert {stmt.bool_expr} under scope {canonize_bounded_variables(last_result.scopes, scope_id)}:', bool(rv.item())) if bool(rv.item()): return [last_result] else: return [] else: raise ValueError(f'Unknown statement type: {stmt}')