Vlm pipeline testing

predicators/approaches/base_approach.py

@@ -11,7 +11,7 @@
 from predicators.structs import Action, Dataset, InteractionRequest, \
     InteractionResult, Metrics, ParameterizedOption, Predicate, State, Task, \
     Type
-from predicators.utils import ExceptionWithInfo
+from predicators.utils import ExceptionWithInfo, create_vlm_by_name
 
 
 class BaseApproach(abc.ABC):
@@ -29,6 +29,7 @@ def __init__(self, initial_predicates: Set[Predicate],
         self._train_tasks = train_tasks
         self._metrics: Metrics = defaultdict(float)
         self._set_seed(CFG.seed)
+        self._vlm = create_vlm_by_name(CFG.vlm_model_name)  # pragma: no cover
 
     @classmethod
     @abc.abstractmethod

predicators/approaches/bilevel_planning_approach.py

@@ -58,12 +58,14 @@ def _solve(self, task: Task, timeout: int) -> Callable[[State], Action]:
         seed = self._seed + self._num_calls
         nsrts = self._get_current_nsrts()
         preds = self._get_current_predicates()
-
+        # utils.abstract(task.init, preds, self._vlm)
+        # utils.abstract(task.init, preds, self._vlm)
         # Run task planning only and then greedily sample and execute in the
         # policy.
         if self._plan_without_sim:
             nsrt_plan, atoms_seq, metrics = self._run_task_plan(
                 task, nsrts, preds, timeout, seed)
+            import pdb; pdb.set_trace()
             self._last_nsrt_plan = nsrt_plan
             self._last_atoms_seq = atoms_seq
             policy = utils.nsrt_plan_to_greedy_policy(nsrt_plan, task.goal,

predicators/approaches/spot_wrapper_approach.py

@@ -55,7 +55,7 @@ def _solve(self, task: Task, timeout: int) -> Callable[[State], Action]:
         def _policy(state: State) -> Action:
             nonlocal base_approach_policy, need_stow
             # If we think that we're done, return the done action.
-            if task.goal_holds(state):
+            if task.goal_holds(state, self._vlm):
                 extra_info = SpotActionExtraInfo("done", [], None, tuple(),
                                                  None, tuple())
                 return utils.create_spot_env_action(extra_info)
@@ -102,7 +102,8 @@ def _policy(state: State) -> Action:
                 self._base_approach_has_control = True
                 # Need to call this once here to fix off-by-one issue.
                 atom_seq = self._base_approach.get_execution_monitoring_info()
-                assert all(a.holds(state) for a in atom_seq[0])
+                # TODO: consider reinstating the line below.
+                # assert all(a.holds(state) for a in atom_seq[0])
             # Use the base policy.
             return base_approach_policy(state)