""" GridWorld Warehouse Navigator Gradio + Gymnasium interactive maze game with RL-style rewards. """ import random from collections import deque import gradio as gr import gymnasium as gym import numpy as np from gymnasium import spaces # ── Constants ──────────────────────────────────────────────────────────────── GRID_SIZE = 10 MAX_STEPS = 200 OBS_DENSITY = 0.20 # Actions UP, RIGHT, DOWN, LEFT = 0, 1, 2, 3 ACTION_DELTAS = {UP: (-1, 0), RIGHT: (0, 1), DOWN: (1, 0), LEFT: (0, -1)} # RL rewards R_GOAL = 100.0 R_CLOSER = 1.0 R_FARTHER = -1.5 R_INVALID = -2.0 R_NEW_CELL = 0.5 R_STEP = -0.1 R_TIMEOUT = -10.0 # SVG rendering CELL = 52 PAD = 6 # ── Gymnasium Environment ───────────────────────────────────────────────────── class GridWorldEnv(gym.Env): """ A simple GridWorld environment. Observation: [agent_x_norm, agent_y_norm, goal_x_norm, goal_y_norm] (float32, 0-1) Actions: 0=UP, 1=RIGHT, 2=DOWN, 3=LEFT """ metadata = {"render_modes": ["rgb_array"]} def __init__(self, grid_size: int = GRID_SIZE, obs_density: float = OBS_DENSITY, max_steps: int = MAX_STEPS): super().__init__() self.grid_size = grid_size self.obs_density = obs_density self.max_steps = max_steps self.action_space = spaces.Discrete(4) self.observation_space = spaces.Box( low=0.0, high=1.0, shape=(4,), dtype=np.float32 ) # State – initialised by reset() self.grid: list[list[str]] = [] self.agent_pos: tuple[int, int] = (0, 0) self.start_pos: tuple[int, int] = (0, 0) self.goal_pos: tuple[int, int] = (0, 0) self.steps = 0 self.total_score = 0.0 self.last_reward = 0.0 self.visited: set[tuple[int, int]] = set() self.goal_reached = False self.reset() # ── maze generation ─────────────────────────────────────────────────────── def _bfs_connected(self, grid, start, goal) -> bool: queue = deque([start]) seen = {start} while queue: r, c = queue.popleft() if (r, c) == goal: return True for dr, dc in ACTION_DELTAS.values(): nr, nc = r + dr, c + dc if 0 <= nr < self.grid_size and 0 <= nc < self.grid_size: if grid[nr][nc] != 'X' and (nr, nc) not in seen: seen.add((nr, nc)) queue.append((nr, nc)) return False def _make_grid(self): """Generate a random, solvable grid. Retries up to 200 times.""" for _ in range(200): g = [['.' for _ in range(self.grid_size)] for _ in range(self.grid_size)] for r in range(self.grid_size): for c in range(self.grid_size): if random.random() < self.obs_density: g[r][c] = 'X' empties = [(r, c) for r in range(self.grid_size) for c in range(self.grid_size) if g[r][c] == '.'] if len(empties) < 2: continue start = random.choice(empties) empties.remove(start) goal = random.choice(empties) if self._bfs_connected(g, start, goal): g[start[0]][start[1]] = 'S' g[goal[0]][goal[1]] = 'G' return g, start, goal # Fallback: empty grid, corners g = [['.' for _ in range(self.grid_size)] for _ in range(self.grid_size)] start, goal = (0, 0), (self.grid_size - 1, self.grid_size - 1) g[start[0]][start[1]] = 'S' g[goal[0]][goal[1]] = 'G' return g, start, goal # ── gym interface ───────────────────────────────────────────────────────── def reset(self, *, seed=None, options=None): super().reset(seed=seed) self.grid, self.start_pos, self.goal_pos = self._make_grid() self.agent_pos = self.start_pos self.steps = 0 self.total_score = 0.0 self.last_reward = 0.0 self.visited = {self.agent_pos} self.goal_reached = False return self._obs(), {} def step(self, action: int): if self.goal_reached or self.steps >= self.max_steps: return self._obs(), 0.0, True, False, {} dr, dc = ACTION_DELTAS[action] nr, nc = self.agent_pos[0] + dr, self.agent_pos[1] + dc prev_dist = self._manhattan(self.agent_pos, self.goal_pos) # ── invalid move ────────────────────────────────────────────────────── if not (0 <= nr < self.grid_size and 0 <= nc < self.grid_size) \ or self.grid[nr][nc] == 'X': self.steps += 1 reward = R_INVALID truncated = self.steps >= self.max_steps if truncated: reward += R_TIMEOUT self.last_reward = reward self.total_score += reward return self._obs(), reward, False, truncated, {"invalid": True} # ── valid move ──────────────────────────────────────────────────────── self.agent_pos = (nr, nc) self.steps += 1 new_dist = self._manhattan(self.agent_pos, self.goal_pos) if self.agent_pos == self.goal_pos: reward = R_GOAL self.goal_reached = True terminated = True truncated = False else: reward = R_STEP reward += R_CLOSER if new_dist < prev_dist else R_FARTHER if self.agent_pos not in self.visited: reward += R_NEW_CELL self.visited.add(self.agent_pos) terminated = False truncated = self.steps >= self.max_steps if truncated: reward += R_TIMEOUT self.last_reward = reward self.total_score += reward return self._obs(), reward, terminated, truncated, {} # ── helpers ─────────────────────────────────────────────────────────────── def _obs(self) -> np.ndarray: n = self.grid_size - 1 ax = self.agent_pos[1] / n ay = self.agent_pos[0] / n gx = self.goal_pos[1] / n gy = self.goal_pos[0] / n return np.array([ax, ay, gx, gy], dtype=np.float32) def _manhattan(self, a, b) -> int: return abs(a[0] - b[0]) + abs(a[1] - b[1]) # ── Single shared environment instance ─────────────────────────────────────── env = GridWorldEnv() # ── SVG grid renderer ───────────────────────────────────────────────────────── def render_svg(e: GridWorldEnv) -> str: n = e.grid_size W = n * CELL + PAD * 2 H = n * CELL + PAD * 2 parts = [ f'') return ''.join(parts) # ── Scoreboard HTML ─────────────────────────────────────────────────────────── def render_scoreboard(e: GridWorldEnv) -> str: dist = e._manhattan(e.agent_pos, e.goal_pos) timeout = e.steps >= e.max_steps and not e.goal_reached if e.goal_reached: status_text = '✅ GOAL REACHED!' status_color = '#69F0AE' elif timeout: status_text = '⏰ TIME LIMIT' status_color = '#FF6B6B' else: status_text = '🔄 In progress' status_color = '#90CAF9' lr_color = '#69F0AE' if e.last_reward > 0 else ('#FF6B6B' if e.last_reward < 0 else '#aaa') def row(label, value, color='#e0e0ff'): return ( f'
| +100 | Reach goal |
| +1.0 | Move closer |
| +0.5 | Visit new cell |
| -0.1 | Each step |
| -1.5 | Move farther |
| -2.0 | Invalid move |
| -10 | Timeout |