diff --git a/src/overcooked_ai_py/mdp/layout_generator.py b/src/overcooked_ai_py/mdp/layout_generator.py index b87a5a72..80474486 100644 --- a/src/overcooked_ai_py/mdp/layout_generator.py +++ b/src/overcooked_ai_py/mdp/layout_generator.py @@ -1,8 +1,9 @@ import numpy as np -import random + +import random, copy from overcooked_ai_py.utils import rnd_int_uniform, rnd_uniform from overcooked_ai_py.mdp.actions import Action, Direction -from overcooked_ai_py.mdp.overcooked_mdp import OvercookedGridworld +from overcooked_ai_py.mdp.overcooked_mdp import OvercookedGridworld, Recipe EMPTY = ' ' COUNTER = 'X' @@ -81,6 +82,7 @@ def generate(self, outside_information={}): mdp_params = self.params_schedule_fn(outside_information) return mdp_params +DEFAULT_FEATURE_TYPES = (POT, ONION_DISPENSER, DISH_DISPENSER, SERVING_LOC) # NOTE: TOMATO_DISPENSER is disabled by default class LayoutGenerator(object): # NOTE: This class hasn't been tested extensively. @@ -126,36 +128,71 @@ def generate_padded_mdp(self, outside_information={}): Return a PADDED MDP with mdp params specified in self.mdp_params """ mdp_gen_params = self.mdp_params_generator.generate(outside_information) + outer_shape = self.outer_shape if "layout_name" in mdp_gen_params.keys() and mdp_gen_params["layout_name"] is not None: mdp = OvercookedGridworld.from_layout_name(**mdp_gen_params) mdp_generator_fn = lambda: self.padded_mdp(mdp) else: - - required_keys = ["inner_shape", "prop_empty", "prop_feats", "display", "start_all_orders"] + required_keys = ["inner_shape", "prop_empty", "prop_feats", "display"] + # with generate_all_orders key start_all_orders will be generated inside make_new_layout method + if not mdp_gen_params.get("generate_all_orders"): + required_keys.append("start_all_orders") missing_keys = [k for k in required_keys if k not in mdp_gen_params.keys()] + if len(missing_keys) != 0: + print("missing keys dict", mdp_gen_params) assert len(missing_keys) == 0, "These keys were missing from the mdp_params: {}".format(missing_keys) - - recipe_params = {"start_all_orders": mdp_gen_params["start_all_orders"]} - if "recipe_values" in mdp_gen_params: - recipe_params["recipe_values"] = mdp_gen_params["recipe_values"] - if "recipe_times" in mdp_gen_params: - recipe_params["recipe_times"] = mdp_gen_params["recipe_times"] - inner_shape = mdp_gen_params["inner_shape"] assert inner_shape[0] <= outer_shape[0] and inner_shape[1] <= outer_shape[1], \ "inner_shape cannot fit into the outershap" - layout_generator = LayoutGenerator(self.mdp_params_generator, outer_shape=self.outer_shape) - mdp_generator_fn = lambda: layout_generator.make_disjoint_sets_layout( - inner_shape=mdp_gen_params["inner_shape"], - prop_empty=mdp_gen_params["prop_empty"], - prop_features=mdp_gen_params["prop_feats"], - base_param=recipe_params, - display=mdp_gen_params["display"] - ) - + + if "feature_types" not in mdp_gen_params: + mdp_gen_params["feature_types"] = DEFAULT_FEATURE_TYPES + + mdp_generator_fn = lambda: layout_generator.make_new_layout(mdp_gen_params) return mdp_generator_fn() + + @staticmethod + def create_base_params(mdp_gen_params): + assert mdp_gen_params.get("start_all_orders") or mdp_gen_params.get("generate_all_orders") + mdp_gen_params = LayoutGenerator.add_generated_mdp_params_orders(mdp_gen_params) + recipe_params = {"start_all_orders": mdp_gen_params["start_all_orders"]} + if mdp_gen_params.get("start_bonus_orders"): + recipe_params["start_bonus_orders"] = mdp_gen_params["start_bonus_orders"] + if "recipe_values" in mdp_gen_params: + recipe_params["recipe_values"] = mdp_gen_params["recipe_values"] + if "recipe_times" in mdp_gen_params: + recipe_params["recipe_times"] = mdp_gen_params["recipe_times"] + return recipe_params + + @staticmethod + def add_generated_mdp_params_orders(mdp_params): + """ + adds generated parameters (i.e. generated orders) to mdp_params, + returns onchanged copy of mdp_params when there is no "generate_all_orders" and "generate_bonus_orders" keys inside mdp_params + """ + mdp_params = copy.deepcopy(mdp_params) + if mdp_params.get("generate_all_orders"): + all_orders_kwargs = copy.deepcopy(mdp_params["generate_all_orders"]) + + if all_orders_kwargs.get("recipes"): + all_orders_kwargs["recipes"] = [Recipe.from_dict(r) for r in all_orders_kwargs["recipes"]] + + all_recipes = Recipe.generate_random_recipes(**all_orders_kwargs) + mdp_params["start_all_orders"] = [r.to_dict() for r in all_recipes] + else: + all_recipes = Recipe.ALL_RECIPES + + if mdp_params.get("generate_bonus_orders"): + bonus_orders_kwargs = copy.deepcopy(mdp_params["generate_bonus_orders"]) + + if not bonus_orders_kwargs.get("recipes"): + bonus_orders_kwargs["recipes"] = all_recipes + + bonus_recipes = Recipe.generate_random_recipes(**bonus_orders_kwargs) + mdp_params["start_bonus_orders"] = [r.to_dict() for r in bonus_recipes] + return mdp_params def padded_mdp(self, mdp, display=False): """Returns a padded MDP from an MDP""" @@ -166,10 +203,20 @@ def padded_mdp(self, mdp, display=False): mdp_grid = self.padded_grid_to_layout_grid(padded_grid, start_positions, display=display) return OvercookedGridworld.from_grid(mdp_grid) - def make_disjoint_sets_layout(self, inner_shape, prop_empty, prop_features, base_param={}, display=True): + def make_new_layout(self, mdp_gen_params): + return self.make_disjoint_sets_layout( + inner_shape=mdp_gen_params["inner_shape"], + prop_empty=mdp_gen_params["prop_empty"], + prop_features=mdp_gen_params["prop_feats"], + base_param=LayoutGenerator.create_base_params(mdp_gen_params), + feature_types=mdp_gen_params["feature_types"], + display=mdp_gen_params["display"] + ) + + def make_disjoint_sets_layout(self, inner_shape, prop_empty, prop_features, base_param, feature_types=DEFAULT_FEATURE_TYPES, display=True): grid = Grid(inner_shape) self.dig_space_with_disjoint_sets(grid, prop_empty) - self.add_features(grid, prop_features) + self.add_features(grid, prop_features, feature_types) padded_grid = self.embed_grid(grid) start_positions = self.get_random_starting_positions(padded_grid) @@ -242,12 +289,10 @@ def dig_space_with_fringe_expansion(self, grid, prop_empty=0.1): if grid.is_valid_dig_location(location): fringe.add(location) - def add_features(self, grid, prop_features=0): + def add_features(self, grid, prop_features=0, feature_types=DEFAULT_FEATURE_TYPES): """ Places one round of basic features and then adds random features until prop_features of valid locations are filled""" - feature_types = [POT, ONION_DISPENSER, DISH_DISPENSER, SERVING_LOC] - # feature_types = [POT, ONION_DISPENSER, TOMATO_DISPENSER, DISH_DISPENSER, SERVING_LOC] # NOTE: currently disabled TOMATO_DISPENSER valid_locations = grid.valid_feature_locations() np.random.shuffle(valid_locations) diff --git a/src/overcooked_ai_py/mdp/overcooked_mdp.py b/src/overcooked_ai_py/mdp/overcooked_mdp.py index afd96118..7b98adf8 100644 --- a/src/overcooked_ai_py/mdp/overcooked_mdp.py +++ b/src/overcooked_ai_py/mdp/overcooked_mdp.py @@ -236,6 +236,34 @@ def configure(cls, conf): if 'onion_value' in conf: cls._onion_value = conf['onion_value'] + + @classmethod + def generate_random_recipes(cls, n=1, min_size=2, max_size=3, ingredients=None, recipes=None, unique=True): + """ + n (int): how many recipes generate + min_size (int): min generated recipe size + max_size (int): max generated recipe size + ingredients (list(str)): list of ingredients used for generating recipes (default is cls.ALL_INGREDIENTS) + recipes (list(Recipe)): list of recipes to choose from (default is cls.ALL_RECIPES) + unique (bool): if all recipes are unique (without repeats) + """ + if recipes is None: recipes = cls.ALL_RECIPES + + ingredients = set(ingredients or cls.ALL_INGREDIENTS) + choice_replace = not(unique) + + assert 1 <= min_size <= max_size <= cls.MAX_NUM_INGREDIENTS + assert all(ingredient in cls.ALL_INGREDIENTS for ingredient in ingredients) + + def valid_size(r): + return min_size <= len(r.ingredients) <= max_size + + def valid_ingredients(r): + return all(i in ingredients for i in r.ingredients) + + relevant_recipes = [r for r in recipes if valid_size(r) and valid_ingredients(r)] + assert choice_replace or (n <= len(relevant_recipes)) + return np.random.choice(relevant_recipes, n, replace=choice_replace) @classmethod def from_dict(cls, obj_dict): diff --git a/testing/overcooked_test.py b/testing/overcooked_test.py index e51946b3..542bc4d4 100644 --- a/testing/overcooked_test.py +++ b/testing/overcooked_test.py @@ -1,17 +1,17 @@ import unittest, os +import json import numpy as np from math import factorial from overcooked_ai_py.mdp.actions import Action, Direction from overcooked_ai_py.mdp.overcooked_mdp import PlayerState, OvercookedGridworld, OvercookedState, ObjectState, SoupState, Recipe from overcooked_ai_py.mdp.overcooked_env import OvercookedEnv, DEFAULT_ENV_PARAMS -from overcooked_ai_py.mdp.layout_generator import LayoutGenerator +from overcooked_ai_py.mdp.layout_generator import LayoutGenerator, ONION_DISPENSER, TOMATO_DISPENSER, POT, DISH_DISPENSER, SERVING_LOC from overcooked_ai_py.agents.agent import AgentGroup, AgentPair, GreedyHumanModel, FixedPlanAgent, RandomAgent from overcooked_ai_py.agents.benchmarking import AgentEvaluator from overcooked_ai_py.planning.planners import MediumLevelActionManager, NO_COUNTERS_PARAMS, MotionPlanner from overcooked_ai_py.utils import save_pickle, load_pickle, iterate_over_json_files_in_dir, load_from_json, save_as_json from utils import TESTING_DATA_DIR, generate_serialized_trajectory - START_ORDER_LIST = ["any"] n, s = Direction.NORTH, Direction.SOUTH e, w = Direction.EAST, Direction.WEST @@ -80,6 +80,24 @@ def test_invalid_input(self): self.assertRaises(ValueError, Recipe, []) self.assertRaises(ValueError, Recipe, "invalid argument") + def test_recipes_generation(self): + self.assertRaises(AssertionError, Recipe.generate_random_recipes, max_size=Recipe.MAX_NUM_INGREDIENTS+1) + self.assertRaises(AssertionError, Recipe.generate_random_recipes, min_size=0) + self.assertRaises(AssertionError, Recipe.generate_random_recipes, min_size=3, max_size=2) + self.assertRaises(AssertionError, Recipe.generate_random_recipes, ingredients=["onion", "tomato", "fake_ingredient"]) + self.assertRaises(AssertionError, Recipe.generate_random_recipes, n=99999) + self.assertEqual(len(Recipe.generate_random_recipes(n=3)), 3) + self.assertEqual(len(Recipe.generate_random_recipes(n=99, unique=False)), 99) + + two_sized_recipes = [Recipe(["onion", "onion"]), Recipe(["onion", "tomato"]), Recipe(["tomato", "tomato"])] + for _ in range(100): + self.assertCountEqual(two_sized_recipes, Recipe.generate_random_recipes(n=3, min_size=2, max_size=2, ingredients=["onion", "tomato"])) + + only_onions_recipes = [Recipe(["onion", "onion"]), Recipe(["onion", "onion", "onion"])] + for _ in range(100): + self.assertCountEqual(only_onions_recipes, Recipe.generate_random_recipes(n=2, min_size=2, max_size=3, ingredients=["onion"])) + + self.assertCountEqual(only_onions_recipes, set([Recipe.generate_random_recipes(n=1, recipes=only_onions_recipes)[0] for _ in range(100)])) # false positives rate for this test is 1/10^99 def _expected_num_recipes(self, num_ingredients, max_len): return comb(num_ingredients + max_len, num_ingredients) - 1 @@ -873,6 +891,81 @@ def test_random_layout(self): all_same_layout = all([np.array_equal(env.mdp.terrain_mtx, terrain) for terrain in layouts_seen]) self.assertFalse(all_same_layout) + def test_random_layout_feature_types(self): + mandatory_features = {POT, DISH_DISPENSER, SERVING_LOC} + optional_features = {ONION_DISPENSER, TOMATO_DISPENSER} + optional_features_combinations = [{ONION_DISPENSER, TOMATO_DISPENSER}, {ONION_DISPENSER}, {TOMATO_DISPENSER}] + + for optional_features_combo in optional_features_combinations: + left_out_optional_features = optional_features - optional_features_combo + used_features = list(optional_features_combo | mandatory_features) + mdp_gen_params = {"prop_feats": 0.9, + "feature_types": used_features, + "prop_empty": 0.1, + "inner_shape": (6, 5), + "display": False, + "start_all_orders" : [ + { "ingredients" : ["onion", "onion", "onion"]} + ]} + mdp_fn = LayoutGenerator.mdp_gen_fn_from_dict(mdp_gen_params, outer_shape=(6, 5)) + env = OvercookedEnv(mdp_fn, **DEFAULT_ENV_PARAMS) + for _ in range(10): + env.reset() + curr_terrain = env.mdp.terrain_mtx + terrain_features = set.union(*(set(line) for line in curr_terrain)) + self.assertTrue(all(elem in terrain_features for elem in used_features)) # all used_features are actually used + if left_out_optional_features: + self.assertFalse(any(elem in terrain_features for elem in left_out_optional_features)) # all left_out optional_features are not used + + def test_random_layout_generated_recipes(self): + only_onions_recipes = [Recipe(["onion", "onion"]), Recipe(["onion", "onion", "onion"])] + only_onions_dict_recipes = [r.to_dict() for r in only_onions_recipes] + + # checking if recipes are generated from mdp_params + mdp_gen_params = {"generate_all_orders": {"n":2, "ingredients": ["onion"], "min_size":2, "max_size":3}, + "prop_feats": 0.9, + "prop_empty": 0.1, + "inner_shape": (6, 5), + "display": False} + mdp_fn = LayoutGenerator.mdp_gen_fn_from_dict(mdp_gen_params, outer_shape=(6, 5)) + env = OvercookedEnv(mdp_fn, **DEFAULT_ENV_PARAMS) + for _ in range(10): + env.reset() + self.assertCountEqual(env.mdp.start_all_orders, only_onions_dict_recipes) + self.assertEqual(len(env.mdp.start_bonus_orders), 0) + + # checking if bonus_orders is subset of all_orders even if not specified + + mdp_gen_params = {"generate_all_orders": {"n":2, "ingredients": ["onion"], "min_size":2, "max_size":3}, + "generate_bonus_orders": {"n":1, "min_size":2, "max_size":3}, + "prop_feats": 0.9, + "prop_empty": 0.1, + "inner_shape": (6, 5), + "display": False} + mdp_fn = LayoutGenerator.mdp_gen_fn_from_dict(mdp_gen_params, outer_shape=(6,5)) + env = OvercookedEnv(mdp_fn, **DEFAULT_ENV_PARAMS) + for _ in range(10): + env.reset() + self.assertCountEqual(env.mdp.start_all_orders, only_onions_dict_recipes) + self.assertEqual(len(env.mdp.start_bonus_orders), 1) + self.assertTrue(env.mdp.start_bonus_orders[0] in only_onions_dict_recipes) + + # checking if after reset there are new recipes generated + mdp_gen_params = {"generate_all_orders": {"n":3, "min_size":2, "max_size":3}, + "prop_feats": 0.9, + "prop_empty": 0.1, + "inner_shape": (6, 5), + "display": False, + "feature_types": [POT, DISH_DISPENSER, SERVING_LOC, ONION_DISPENSER, TOMATO_DISPENSER] + } + mdp_fn = LayoutGenerator.mdp_gen_fn_from_dict(mdp_gen_params, outer_shape=(6,5)) + env = OvercookedEnv(mdp_fn, **DEFAULT_ENV_PARAMS) + generated_recipes_strings = set() + for _ in range(20): + env.reset() + generated_recipes_strings |= {json.dumps(o, sort_keys=True) for o in env.mdp.start_all_orders} + self.assertTrue(len(generated_recipes_strings) > 3) + class TestGymEnvironment(unittest.TestCase):