run_tests.py 2.8 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100
  1. #!/usr/bin/env python3
  2. import random
  3. import numpy as np
  4. import gym
  5. from gym_minigrid.register import env_list
  6. from gym_minigrid.minigrid import Grid, OBJECT_TO_IDX
  7. # Test specifically importing a specific environment
  8. from gym_minigrid.envs import DoorKeyEnv
  9. # Test importing wrappers
  10. from gym_minigrid.wrappers import *
  11. ##############################################################################
  12. print('%d environments registered' % len(env_list))
  13. for envName in env_list:
  14. print('testing "%s"' % envName)
  15. # Load the gym environment
  16. env = gym.make(envName)
  17. env.max_steps = min(env.max_steps, 200)
  18. env.reset()
  19. env.render('rgb_array')
  20. # Verify that the same seed always produces the same environment
  21. for i in range(0, 5):
  22. seed = 1337 + i
  23. env.seed(seed)
  24. grid1 = env.grid
  25. env.seed(seed)
  26. grid2 = env.grid
  27. assert grid1 == grid2
  28. env.reset()
  29. # Run for a few episodes
  30. num_episodes = 0
  31. while num_episodes < 5:
  32. # Pick a random action
  33. action = random.randint(0, env.action_space.n - 1)
  34. obs, reward, done, info = env.step(action)
  35. # Validate the agent position
  36. assert env.agent_pos[0] < env.width
  37. assert env.agent_pos[1] < env.height
  38. # Test observation encode/decode roundtrip
  39. img = obs['image']
  40. vis_mask = img[:, :, 0] != OBJECT_TO_IDX['unseen'] # hackish
  41. img2 = Grid.decode(img).encode(vis_mask=vis_mask)
  42. assert np.array_equal(img, img2)
  43. # Test the env to string function
  44. str(env)
  45. # Check that the reward is within the specified range
  46. assert reward >= env.reward_range[0], reward
  47. assert reward <= env.reward_range[1], reward
  48. if done:
  49. num_episodes += 1
  50. env.reset()
  51. env.render('rgb_array')
  52. # Test the fully observable wrapper
  53. env = FullyObsWrapper(env)
  54. env.reset()
  55. obs, _, _, _ = env.step(0)
  56. assert obs.shape == env.observation_space.shape
  57. env.close()
  58. ##############################################################################
  59. print('testing agent_sees method')
  60. env = gym.make('MiniGrid-DoorKey-6x6-v0')
  61. goal_pos = (env.grid.width - 2, env.grid.height - 2)
  62. # Test the "in" operator on grid objects
  63. assert ('green', 'goal') in env.grid
  64. assert ('blue', 'key') not in env.grid
  65. # Test the env.agent_sees() function
  66. env.reset()
  67. for i in range(0, 500):
  68. action = random.randint(0, env.action_space.n - 1)
  69. obs, reward, done, info = env.step(action)
  70. goal_visible = ('green', 'goal') in Grid.decode(obs['image'])
  71. agent_sees_goal = env.agent_sees(*goal_pos)
  72. assert agent_sees_goal == goal_visible
  73. if done:
  74. env.reset()
  75. #############################################################################