synth-ai 0.2.4.dev4__py3-none-any.whl → 0.2.4.dev5__py3-none-any.whl

synth_ai/environments/examples/sokoban/engine_helpers/room_utils.py

@@ -0,0 +1,656 @@
+import numpy as np
+import marshal
+import copy
+from collections import deque
+from typing import Tuple
+
+# --- ADDED CONSTANT ---
+MAX_GENERATION_RETRIES = 100
+# --- END ADDED CONSTANT ---
+
+# import matplotlib.pyplot as plt
+# import matplotlib.animation as animation
+
+
+def get_shortest_action_path(room_fixed, room_state, MAX_DEPTH=100):
+    """
+    Get the shortest action path to push all boxes to the target spots.
+    Use BFS to find the shortest path.
+    NOTE currently only support one player, only one shortest solution
+    =========================================================
+    Parameters:
+        room_state (np.ndarray): the state of the room
+            - 0: wall
+            - 1: empty space
+            - 2: box target
+            - 3: box on target
+            - 4: box not on target
+            - 5: player
+        room_fixed (np.ndarray): the fixed part of the room
+            - 0: wall
+            - 1: empty space
+            - 2: box target
+        MAX_DEPTH (int): the maximum depth of the search
+    =========================================================
+    Returns:
+        action_sequence (list): the action sequence to push all boxes to the target spots
+    """
+
+    # BFS queue stores (room_state, path)
+    queue = deque([(copy.deepcopy(room_state), [])])
+    explored_states = set()
+
+    # Possible moves: up, down, left, right
+    moves = [(-1, 0), (1, 0), (0, -1), (0, 1)]
+    actions = [1, 2, 3, 4]  # Corresponding action numbers
+
+    while queue:
+        room_state, path = queue.popleft()
+        if len(path) > MAX_DEPTH:
+            return []  # No solution found
+
+        # reduce the search space by checking if the state has been explored
+        state_tohash = marshal.dumps(room_state)
+        if state_tohash in explored_states:
+            continue
+        explored_states.add(state_tohash)
+
+        # get information of the room
+        player_pos = tuple(np.argwhere(room_state == 5)[0])
+        boxes_on_target = set(map(tuple, np.argwhere((room_state == 3))))
+        boxes_not_on_target = set(map(tuple, np.argwhere((room_state == 4))))
+        boxes = boxes_on_target | boxes_not_on_target
+
+        # Check if all boxes are on targets
+        if not boxes_not_on_target:
+            return path
+
+        # Try each direction
+        for move, action in zip(moves, actions):
+            new_room_state = copy.deepcopy(room_state)
+            new_player_pos = (player_pos[0] + move[0], player_pos[1] + move[1])
+
+            # Check is new player position is wall or out of bound
+            if (
+                new_player_pos[0] < 0
+                or new_player_pos[0] >= room_fixed.shape[0]
+                or new_player_pos[1] < 0
+                or new_player_pos[1] >= room_fixed.shape[1]
+                or room_fixed[new_player_pos] == 0
+            ):
+                continue
+
+            # If there's a box, check if we can push it
+            if new_player_pos in boxes:
+                box_pos = new_player_pos  # the original box position
+                new_box_pos = (new_player_pos[0] + move[0], new_player_pos[1] + move[1])
+
+                # Can't push if hitting wall or another box or out of bound
+                if (
+                    room_fixed[new_box_pos] == 0
+                    or new_box_pos in boxes
+                    or new_box_pos[0] < 0
+                    or new_box_pos[0] >= room_fixed.shape[0]
+                    or new_box_pos[1] < 0
+                    or new_box_pos[1] >= room_fixed.shape[1]
+                ):
+                    continue
+
+                # move the box
+
+                new_room_state[box_pos] = room_fixed[box_pos]
+                if room_fixed[new_box_pos] == 2:
+                    new_room_state[new_box_pos] = 3
+                else:
+                    new_room_state[new_box_pos] = 4
+
+            # player moves
+            new_room_state[player_pos] = room_fixed[player_pos]
+            new_room_state[new_player_pos] = 5
+            queue.append((new_room_state, path + [action]))
+
+    return []  # No solution found
+
+
+# def plot_animation(imgs):
+#     fig, ax = plt.subplots()
+#     im = ax.imshow(imgs[0])
+#     def init():
+#         im.set_data(imgs[0])
+#         return [im]
+#     def update(i):
+#         im.set_data(imgs[i])
+#         return [im]
+#     ani = animation.FuncAnimation(fig, update, frames=len(imgs), init_func=init, blit=True)
+#     return ani
+
+
+# def plot_animation(imgs):
+#     height, width = imgs[0].shape[:2]
+#     fig = plt.figure(figsize=(width / 100, height / 100), dpi=500)
+
+#     ax = fig.add_axes([0, 0, 1, 1])
+
+#     ax.set_xticks([])
+#     ax.set_yticks([])
+#     ax.set_frame_on(False)
+
+#     im = ax.imshow(imgs[0])
+
+#     def init():
+#         im.set_data(imgs[0])
+#         return [im]
+
+#     def update(i):
+#         im.set_data(imgs[i])
+#         return [im]
+
+#     ani = animation.FuncAnimation(
+#         fig, update, frames=len(imgs), init_func=init, blit=True
+#     )
+#     return ani
+
+
+# def solve_sokoban(env, saved_animation_path):
+#     """
+#     Solve the given sokoban environment and save the animation
+#     """
+#     actions = get_shortest_action_path(env.room_fixed, env.room_state)
+#     print(f"Found {len(actions)} actions: {actions}")
+#     imgs = []
+#     img_before_action = env.render("rgb_array")
+#     imgs.append(img_before_action)
+#     for action in actions:
+#         env.step(action)
+#         img_after_action = env.render("rgb_array")
+#         imgs.append(img_after_action)
+#     ani = plot_animation(imgs)
+#     ani.save(saved_animation_path)
+
+
+def add_random_player_movement(
+    room_state: np.ndarray,
+    room_structure: np.ndarray,
+    rng: np.random.Generator,
+    move_probability=0.5,
+    continue_probability=0.5,
+    max_steps=3,
+) -> np.ndarray:
+    """
+    Randomly move the player after reverse_playing.
+    Uses the provided RNG instance.
+    """
+    # Check if we should move the player at all
+    if rng.random() > move_probability:
+        return room_state
+
+    # Find player position
+    player_pos_coords = np.where(room_state == 5)
+    if len(player_pos_coords[0]) == 0:
+        # Should not happen if called correctly, but handle gracefully
+        print("Warning: Player not found in room_state during add_random_player_movement.")
+        return room_state
+    player_pos = np.array([player_pos_coords[0][0], player_pos_coords[1][0]])
+
+    # Keep track of previous positions to avoid moving back
+    previous_positions = {tuple(player_pos)}
+
+    # Make 1-3 random moves
+    steps_taken = 0
+    while steps_taken < max_steps:
+        # Get all valid moves (can't move into walls or boxes)
+        valid_moves = []
+        for action in range(4):  # 0: up, 1: down, 2: left, 3: right
+            change = CHANGE_COORDINATES[action]
+            next_pos = player_pos + change
+
+            # Check bounds first
+            if not (
+                0 <= next_pos[0] < room_state.shape[0] and 0 <= next_pos[1] < room_state.shape[1]
+            ):
+                continue
+
+            # Check if next position is valid (empty space or target) and not a previous position
+            if (
+                room_state[next_pos[0], next_pos[1]] in [1, 2]
+                and tuple(next_pos) not in previous_positions
+            ):
+                valid_moves.append((action, next_pos))
+
+        # If no valid moves, break
+        if not valid_moves:
+            break
+
+        # Choose a random valid move using the passed RNG
+        move_index = rng.integers(0, len(valid_moves))
+        chosen_action, next_pos = valid_moves[move_index]
+
+        # Move player
+        room_state[player_pos[0], player_pos[1]] = room_structure[player_pos[0], player_pos[1]]
+        room_state[next_pos[0], next_pos[1]] = 5
+
+        # Update player position and track previous position
+        player_pos = next_pos
+        previous_positions.add(tuple(player_pos))
+
+        steps_taken += 1
+
+        # Decide whether to continue moving using passed RNG
+        if steps_taken >= max_steps or rng.random() > continue_probability:
+            break
+
+    return room_state
+
+
+"""
+Following code is adapted from the nicely written gym_sokoban repo
+"""
+
+
+def generate_room(
+    dim: Tuple[int, int] = (13, 13),
+    initial_seed: int = 42,
+    p_change_directions=0.35,
+    num_steps=25,
+    num_boxes=3,
+    second_player=False,
+    search_depth=100,
+):
+    """
+    Generates a Sokoban room deterministically based on initial_seed.
+    Retries generation with derived seeds if the first attempt fails (score=0).
+    Uses numpy.random.default_rng for internal randomness.
+    """
+
+    for retry_offset in range(MAX_GENERATION_RETRIES):
+        current_seed = initial_seed + retry_offset
+        # Create a new RNG instance for this attempt, seeded deterministically
+        rng = np.random.default_rng(current_seed)
+
+        # Generate topology using the temporary RNG
+        room = room_topology_generation(dim, p_change_directions, num_steps, rng)
+
+        # Place boxes and player using the temporary RNG
+        try:
+            room = place_boxes_and_player(room, num_boxes, second_player, rng)
+        except RuntimeError as e:
+            # This can happen if not enough space; treat as generation failure for this seed
+            print(f"Warning: place_boxes_and_player failed for seed {current_seed}: {e}")
+            continue  # Try next seed offset
+
+        # Prepare for reverse playing
+        room_structure = np.copy(room)
+        room_structure[room_structure == 5] = 1
+        room_state = room.copy()
+        room_state[room_state == 2] = 4  # Mark initial boxes as 'not on target'
+
+        # Perform reverse playing (doesn't need RNG currently)
+        room_state_rev, box_mapping, action_sequence = reverse_playing(
+            room_state, room_structure, search_depth
+        )
+
+        # If reverse playing failed (returned None), try next seed offset
+        if room_state_rev is None:
+            print(f"Warning: reverse_playing failed for seed {current_seed}.")
+            continue
+
+        # Check score
+        score = box_displacement_score(box_mapping)
+        if score > 0:
+            # Successfully generated a potentially solvable room
+            final_room_state = room_state_rev.copy()
+            final_room_state[final_room_state == 3] = (
+                4  # Ensure boxes are marked correctly after reverse play
+            )
+
+            # Add random player movement using the temporary RNG
+            move_prob = 0.8 if score == 1 else 0.5
+            final_room_state = add_random_player_movement(
+                final_room_state,
+                room_structure,
+                rng,  # Pass the temporary RNG
+                move_probability=move_prob,
+                continue_probability=0.5,
+                max_steps=3,
+            )
+            # print(
+            #     f"Successfully generated room with score {score} using seed {current_seed} (offset {retry_offset})."
+            # )
+            return room_structure, final_room_state, box_mapping, action_sequence
+        else:
+            # Score is 0, log and continue to next retry offset
+
+            pass
+            # print(
+            #     f"Warning: Generated room with score 0 for seed {current_seed} (offset {retry_offset}). Retrying..."
+            # )
+
+    # If loop completes without success
+    raise RuntimeWarning(
+        f"Generated Model with score == 0 after {MAX_GENERATION_RETRIES} attempts. "
+        f"Parameters: initial_seed={initial_seed}, dim={dim}, "
+        f"p_change_directions={p_change_directions}, num_steps={num_steps}, "
+        f"num_boxes={num_boxes}, second_player={second_player}, "
+        f"search_depth={search_depth}."
+    )
+
+
+def room_topology_generation(
+    dim: Tuple[int, int],
+    p_change_directions: float,
+    num_steps: int,
+    rng: np.random.Generator,
+) -> np.ndarray:
+    """Generate room topology using the provided RNG instance."""
+    dim_x, dim_y = dim
+
+    # The ones in the mask represent all fields which will be set to floors
+    # during the random walk. The centered one will be placed over the current
+    # position of the walk.
+    masks = [
+        [[0, 0, 0], [1, 1, 1], [0, 0, 0]],
+        [[0, 1, 0], [0, 1, 0], [0, 1, 0]],
+        [[0, 0, 0], [1, 1, 0], [0, 1, 0]],
+        [[0, 0, 0], [1, 1, 0], [1, 1, 0]],
+        [[0, 0, 0], [0, 1, 1], [0, 1, 0]],
+    ]
+
+    directions = [(1, 0), (0, 1), (-1, 0), (0, -1)]
+    # Use rng.choice for single sample
+    direction_index = rng.choice(len(directions))
+    direction = directions[direction_index]
+
+    # Starting position uses rng.integers
+    position = np.array([rng.integers(1, dim_x - 1), rng.integers(1, dim_y - 1)])
+
+    level = np.zeros(dim, dtype=int)
+
+    for _ in range(num_steps):
+        # Change direction randomly using rng.random
+        if rng.random() < p_change_directions:
+            direction_index = rng.choice(len(directions))
+            direction = directions[direction_index]
+
+        # Update position
+        position = position + direction
+        position[0] = max(min(position[0], dim_x - 2), 1)
+        position[1] = max(min(position[1], dim_y - 2), 1)
+
+        # Apply mask using rng.choice
+        mask_index = rng.choice(len(masks))
+        mask = masks[mask_index]
+        mask_start = position - 1
+        level[mask_start[0] : mask_start[0] + 3, mask_start[1] : mask_start[1] + 3] += mask
+
+    level[level > 0] = 1
+    level[:, [0, dim_y - 1]] = 0
+    level[[0, dim_x - 1], :] = 0
+
+    return level
+
+
+def place_boxes_and_player(
+    room: np.ndarray, num_boxes: int, second_player: bool, rng: np.random.Generator
+) -> np.ndarray:
+    """
+    Places player and boxes using the provided RNG instance.
+    Raises RuntimeError if not enough space.
+    """
+    # Get all available positions
+    possible_indices = np.where(room == 1)
+    possible_positions = list(zip(possible_indices[0], possible_indices[1]))
+    num_possible_positions = len(possible_positions)
+    num_players = 2 if second_player else 1
+
+    if num_possible_positions <= num_boxes + num_players:
+        raise RuntimeError(
+            f"Not enough free spots ({num_possible_positions}) to place {num_players} player(s) and {num_boxes} boxes."
+        )
+
+    # Choose positions using rng.choice without replacement
+    chosen_indices = rng.choice(num_possible_positions, size=num_boxes + num_players, replace=False)
+    chosen_positions = [possible_positions[i] for i in chosen_indices]
+
+    # Place player(s)
+    player_positions = chosen_positions[:num_players]
+    for pos in player_positions:
+        room[pos] = 5
+
+    # Place boxes
+    box_positions = chosen_positions[num_players:]
+    for pos in box_positions:
+        room[pos] = 2  # Place as box target initially
+
+    return room
+
+
+# Global variables used for reverse playing.
+explored_states = set()
+num_boxes = 0
+best_room_score = -1
+best_room = None
+best_box_mapping = None
+
+
+def reverse_playing(room_state, room_structure, search_depth=100):
+    """
+    This function plays Sokoban reverse in a way, such that the player can
+    move and pull boxes.
+    It ensures a solvable level with all boxes not being placed on a box target.
+    :param room_state:
+    :param room_structure:
+    :param search_depth:
+    :return: 2d array, box mapping, action sequence
+    """
+    global \
+        explored_states, \
+        num_boxes, \
+        best_room_score, \
+        best_room, \
+        best_box_mapping, \
+        best_action_sequence
+
+    # Box_Mapping is used to calculate the box displacement for every box
+    box_mapping = {}
+    box_locations = np.where(room_structure == 2)
+    num_boxes = len(box_locations[0])
+    for l in range(num_boxes):
+        box = (box_locations[0][l], box_locations[1][l])
+        box_mapping[box] = box
+
+    # explored_states globally stores the best room state and score found during search
+    explored_states = set()
+    best_room_score = -1
+    best_room = None
+    best_box_mapping = box_mapping
+    best_action_sequence = []
+
+    depth_first_search(
+        room_state,
+        room_structure,
+        box_mapping,
+        box_swaps=0,
+        last_pull=(-1, -1),
+        ttl=search_depth,
+        action_sequence=[],
+    )
+
+    return best_room, best_box_mapping, best_action_sequence
+
+
+def depth_first_search(
+    room_state,
+    room_structure,
+    box_mapping,
+    box_swaps=0,
+    last_pull=(-1, -1),
+    ttl=300,
+    action_sequence=[],
+):
+    """
+    Searches through all possible states of the room.
+    This is a recursive function, which stops if the ttl is reduced to 0 or
+    over 1.000.000 states have been explored.
+    :param room_state:
+    :param room_structure:
+    :param box_mapping:
+    :param box_swaps:
+    :param last_pull:
+    :param ttl:
+    :param action_sequence:
+    :return:
+    """
+    global \
+        explored_states, \
+        num_boxes, \
+        best_room_score, \
+        best_room, \
+        best_box_mapping, \
+        best_action_sequence
+
+    ttl -= 1
+    if ttl <= 0 or len(explored_states) >= 300000:
+        return
+
+    state_tohash = marshal.dumps(room_state)
+
+    # Only search this state, if it not yet has been explored
+    if state_tohash not in explored_states:
+        # Add current state and its score to explored states
+        room_score = box_swaps * box_displacement_score(box_mapping)
+        if np.where(room_state == 2)[0].shape[0] != num_boxes:
+            room_score = 0
+
+        if room_score > best_room_score:
+            best_room = room_state.copy()
+            best_room_score = room_score
+            best_box_mapping = box_mapping.copy()
+            best_action_sequence = action_sequence.copy()
+
+        explored_states.add(state_tohash)
+
+        for action in ACTION_LOOKUP.keys():
+            # The state and box mapping need to be copied to ensure
+            # every action starts from a similar state.
+
+            # TODO: A tentitive try here to make less moves
+            if action >= 4:
+                continue
+
+            room_state_next = room_state.copy()
+            box_mapping_next = box_mapping.copy()
+
+            room_state_next, box_mapping_next, last_pull_next = reverse_move(
+                room_state_next, room_structure, box_mapping_next, last_pull, action
+            )
+
+            box_swaps_next = box_swaps
+            if last_pull_next != last_pull:
+                box_swaps_next += 1
+
+            action_sequence_next = action_sequence + [action]
+            # action_sequence_next = action_sequence + [(action, box_mapping_next != box_mapping)] # add whether a box is moved
+            depth_first_search(
+                room_state_next,
+                room_structure,
+                box_mapping_next,
+                box_swaps_next,
+                last_pull_next,
+                ttl,
+                action_sequence_next,
+            )
+
+
+def reverse_move(room_state, room_structure, box_mapping, last_pull, action):
+    """
+    Perform reverse action. Where all actions in the range [0, 3] correspond to
+    push actions and the ones greater 3 are simmple move actions.
+    :param room_state:
+    :param room_structure:
+    :param box_mapping:
+    :param last_pull:
+    :param action:
+    :return:
+    """
+    player_position = np.where(room_state == 5)
+    player_position = np.array([player_position[0][0], player_position[1][0]])
+
+    change = CHANGE_COORDINATES[action % 4]
+    next_position = player_position + change
+
+    # Check if next position is an empty floor or an empty box target
+    if room_state[next_position[0], next_position[1]] in [1, 2]:
+        # Move player, independent of pull or move action.
+        room_state[player_position[0], player_position[1]] = room_structure[
+            player_position[0], player_position[1]
+        ]
+        room_state[next_position[0], next_position[1]] = 5
+
+        # In addition try to pull a box if the action is a pull action
+        if action < 4:
+            possible_box_location = change[0] * -1, change[1] * -1
+            possible_box_location += player_position
+
+            if room_state[possible_box_location[0], possible_box_location[1]] in [3, 4]:
+                # Perform pull of the adjacent box
+                room_state[player_position[0], player_position[1]] = 3
+                room_state[possible_box_location[0], possible_box_location[1]] = room_structure[
+                    possible_box_location[0], possible_box_location[1]
+                ]
+
+                # Update the box mapping
+                for k in box_mapping.keys():
+                    if box_mapping[k] == (
+                        possible_box_location[0],
+                        possible_box_location[1],
+                    ):
+                        box_mapping[k] = (player_position[0], player_position[1])
+                        last_pull = k
+
+    return room_state, box_mapping, last_pull
+
+
+def box_displacement_score(box_mapping):
+    """
+    Calculates the sum of all Manhattan distances, between the boxes
+    and their origin box targets.
+    :param box_mapping:
+    :return:
+    """
+    score = 0
+
+    for box_target in box_mapping.keys():
+        box_location = np.array(box_mapping[box_target])
+        box_target = np.array(box_target)
+        dist = np.sum(np.abs(box_location - box_target))
+        score += dist
+
+    return score
+
+
+TYPE_LOOKUP = {
+    0: "wall",
+    1: "empty space",
+    2: "box target",
+    3: "box on target",
+    4: "box not on target",
+    5: "player",
+}
+
+ACTION_LOOKUP = {
+    0: "push up",
+    1: "push down",
+    2: "push left",
+    3: "push right",
+    4: "move up",
+    5: "move down",
+    6: "move left",
+    7: "move right",
+}
+
+# Moves are mapped to coordinate changes as follows
+# 0: Move up
+# 1: Move down
+# 2: Move left
+# 3: Move right
+CHANGE_COORDINATES = {0: (-1, 0), 1: (1, 0), 2: (0, -1), 3: (0, 1)}

synth_ai/environments/examples/sokoban/engine_helpers/vendored/__init__.py

@@ -0,0 +1,17 @@
+import logging
+import pkg_resources
+import json
+from gym.envs.registration import register
+
+logger = logging.getLogger(__name__)
+
+resource_package = __name__
+env_json = pkg_resources.resource_filename(
+    resource_package, "/".join(("envs", "available_envs.json"))
+)
+
+with open(env_json) as f:
+    envs = json.load(f)
+
+    for env in envs:
+        register(id=env["id"], entry_point=env["entry_point"])

synth_ai/environments/examples/sokoban/engine_helpers/vendored/envs/__init__.py

@@ -0,0 +1,3 @@
+from .sokoban_env import SokobanEnv, ACTION_LOOKUP, CHANGE_COORDINATES
+from . import room_utils
+from .sokoban_env_variations import *