def test_agents_on_same_cell():
agent_1 = Agent()
agent_2 = Agent()
simple_jungle = EmptyJungle(size=11)
simple_jungle.add_agents(agent_1, agent_2)
# Move agent 2 on cell of agent 1
# First rotate
# agent_2 rotates
actions = {agent_2: {Actions.ROTATE: 1}}
simple_jungle.step(actions)
simple_jungle.step(actions)
simple_jungle.step(actions)
# then move forward twice
# agent_2 moves fwd
actions = {agent_2: {Actions.FORWARD: 1}}
simple_jungle.step(actions)
_, rew, _ = simple_jungle.step(actions)
# should be on the same cell, no collision
assert agent_1.position == agent_2.position
assert rew[agent_1] == rew[agent_2] == 0
def test_partial_observability():
agent_1 = Agent(range_observation=4)
agent_2 = Agent(range_observation=4)
simple_jungle = EmptyJungle(size=11)
simple_jungle.add_agents(agent_1, agent_2)
simple_jungle.add_object(ElementsEnv.AGENT_1, (5, 4))
simple_jungle.add_object(ElementsEnv.AGENT_2, (5, 6))
simple_jungle.add_object(ElementsEnv.RIVER, (1, 8))
simple_jungle.add_object(ElementsEnv.RIVER, (2, 9))
simple_jungle.add_object(ElementsEnv.EXIT_WHITE, (3, 1))
actions = {}
obs, rew, done = simple_jungle.step(actions)
assert ElementsEnv.EXIT_WHITE.value in obs[agent_1]
assert ElementsEnv.RIVER.value in obs[agent_2]
simple_jungle.add_object(ElementsEnv.BOULDER, (3, 2))
simple_jungle.add_object(ElementsEnv.TREE, (3, 8))
actions = {}
obs, rew, done = simple_jungle.step(actions)
assert ElementsEnv.EXIT_WHITE.value not in obs[agent_1]
assert ElementsEnv.RIVER.value not in obs[agent_2]
def test_obs_mediumenv():
agent_1 = Agent(range_observation=4)
agent_2 = Agent(range_observation=4)
simple_jungle = EmptyJungle(size=11)
simple_jungle.add_agents(agent_1, agent_2)
simple_jungle.add_object(ElementsEnv.AGENT_1, (5, 4))
simple_jungle.add_object(ElementsEnv.AGENT_2, (5, 6))
simple_jungle.add_object(ElementsEnv.EXIT_DIFFICULT, (1, 9))
simple_jungle.add_object(ElementsEnv.RIVER, (1, 8))
simple_jungle.add_object(ElementsEnv.RIVER, (2, 9))
actions = {}
obs, rew, done = simple_jungle.step(actions)
assert ElementsEnv.RIVER.value in obs[agent_2]
simple_jungle.add_object(ElementsEnv.TREE, (4, 7))
simple_jungle.add_object(ElementsEnv.TREE, (3, 8))
simple_jungle.add_object(ElementsEnv.TREE, (3, 7))
actions = {}
obs, rew, done = simple_jungle.step(actions)
assert agent_2.top_right_obstructed
assert ElementsEnv.RIVER.value not in obs[agent_2]
def run_tree_experiment():
# Agent 1 and 2 move at the same time and collide with a tree.
# Sometimes 1 get the log, sometimes 2 get the log
# . . . . . .
# . 1 T 2 .
# . . . . . .
agent_1 = Agent()
agent_2 = Agent()
simple_jungle = EmptyJungle(size=11)
simple_jungle.add_agents(agent_1, agent_2, random_position=False)
simple_jungle.add_object(ElementsEnv.TREE, (5, 5))
# face the tree
# agent_1 and agent_2 rotate -1
actions = {
agent_1: {
Actions.FORWARD: 0,
Actions.ROTATE: -1,
Actions.CLIMB: 0
},
agent_2: {
Actions.FORWARD: 0,
Actions.ROTATE: -1,
Actions.CLIMB: 0
}
}
simple_jungle.step(actions)
simple_jungle.step(actions)
simple_jungle.step(actions)
# move towards the tree
# agent_1 fwd 1 , agent_2 fwd -1
actions = {
agent_1: {
Actions.FORWARD: 1,
Actions.ROTATE: 0,
Actions.CLIMB: 0
},
agent_2: {
Actions.FORWARD: -1,
Actions.ROTATE: 0,
Actions.CLIMB: 0
}
}
simple_jungle.step(actions)
# one of them gets the log
assert ((agent_1.wood_logs == 1 and agent_2.wood_logs == 0)
or (agent_1.wood_logs == 0 and agent_2.wood_logs == 1))
return agent_1.wood_logs, agent_2.wood_logs
def test_obstacles_in_obs_cross():
agent_1 = Agent(range_observation=4)
agent_2 = Agent(range_observation=4)
simple_jungle = EmptyJungle(size=11)
simple_jungle.add_agents(agent_1, agent_2)
# directly left of agent 1
simple_jungle.add_object(ElementsEnv.TREE, (5, 3))
# directly right of agent 2
simple_jungle.add_object(ElementsEnv.TREE, (5, 8))
actions = {
agent_1: {
Actions.FORWARD: 0,
Actions.ROTATE: 0,
Actions.CLIMB: 0
},
agent_2: {
Actions.FORWARD: 0,
Actions.ROTATE: 0,
Actions.CLIMB: 0
}
}
obs, rew, done = simple_jungle.step(actions)
# @MG these assert statements were just to do some checks - they
# will not be part of the final codebase
assert agent_1.left_view_obstructed
assert agent_2.right_view_obstructed
# directly below agent 1
simple_jungle.add_object(ElementsEnv.TREE, (8, 4))
# directly above agent 2
simple_jungle.add_object(ElementsEnv.TREE, (3, 7))
actions = {
agent_1: {
Actions.FORWARD: 1,
Actions.ROTATE: 1,
Actions.CLIMB: 0
},
agent_2: {
Actions.FORWARD: 1,
Actions.ROTATE: 1,
Actions.CLIMB: 0
}
}
obs, rew, done = simple_jungle.step(actions)
assert agent_1.bottom_view_obstructed
assert agent_2.top_view_obstructed
def test_rl_loop():
"""
Tests the general RL api.
"""
agent_1 = Agent()
agent_2 = Agent()
simple_jungle = EmptyJungle(size=11)
assert simple_jungle.size == 11
simple_jungle.add_agents(agent_1, agent_2)
# Once added, each agent randomly takes a color for the game
assert (agent_1.color is BLACK and agent_2.color is WHITE) \
or (agent_1.color is WHITE and agent_2.color is BLACK)
actions = {
agent_1: {
Actions.FORWARD: 1,
Actions.ROTATE: -1
},
agent_2: {
Actions.FORWARD: 1
}
}
obs, rew, done = simple_jungle.step(actions)
# Because observations are not set, it should return none:
assert isinstance(obs[agent_1], dict)
assert isinstance(obs[agent_2], dict)
assert isinstance(rew[agent_1], (float, int))
assert isinstance(rew[agent_2], (float, int))
assert not done[agent_1] and not done[agent_2]
# should work with some actions not set (default to 0
# agent_1 moves fwd ; agent_2 rotates -1
actions = {agent_1: {Actions.FORWARD: 1}, agent_2: {Actions.ROTATE: -1}}
simple_jungle.step(actions)
# should work without agent in the dict. (default to 0)
pos_before = agent_1.position
# agent_2 rotates -1
actions = {agent_2: {Actions.ROTATE: -1}}
simple_jungle.step(actions)
# make sure that position doesn't change when action is empty
assert pos_before == agent_1.position
def test_run_check():
agent_1 = Agent()
agent_2 = Agent()
agents = [agent_1, agent_2]
actions = {}
for agent in agents:
actions[agent] = generate_actions()
print(actions)
return actions
def test_build_bridge():
agent_1 = Agent()
agent_2 = Agent()
simple_jungle = EmptyJungle(size=11)
simple_jungle.add_agents(agent_1, agent_2, random_position=True)
# Place a river between them, and trees in front of them
simple_jungle.add_object(ElementsEnv.RIVER, (5, 5))
simple_jungle.add_object(ElementsEnv.TREE, (5, 3))
simple_jungle.add_object(ElementsEnv.TREE, (5, 2))
simple_jungle.add_object(ElementsEnv.TREE, (5, 7))
simple_jungle.add_object(ElementsEnv.TREE, (5, 8))
# Go fetch some wood
actions = {agent_1: {Actions.FORWARD: 1}, agent_2: {Actions.FORWARD: 1}}
simple_jungle.step(actions)
simple_jungle.step(actions)
# assert there are enough logs
assert (agent_1.wood_logs + agent_2.wood_logs) == 4
# go back to middle
actions = {
agent_1: {
Actions.FORWARD: 0,
Actions.ROTATE: 1
},
agent_2: {
Actions.FORWARD: 0,
Actions.ROTATE: 1
}
}
simple_jungle.step(actions)
simple_jungle.step(actions)
simple_jungle.step(actions)
actions = {agent_1: {Actions.FORWARD: 1}, agent_2: {Actions.FORWARD: 1}}
simple_jungle.step(actions)
simple_jungle.step(actions)
_, rew, done = simple_jungle.step(actions)
# assert river cell becomes empty cell
assert simple_jungle.cell_type(5, 5) == ElementsEnv.EMPTY.value
# assert they both are not done
assert not agent_1.done
assert not agent_2.done
def test_climb_boulders():
# . . B 1 . . 2 .
agent_1 = Agent(range_observation=4)
agent_2 = Agent(range_observation=4)
simple_jungle = EmptyJungle(size=11)
simple_jungle.add_agents(agent_1, agent_2)
actions = {agent_2: {Actions.ROTATE: -1}}
simple_jungle.step(actions)
simple_jungle.step(actions)
simple_jungle.step(actions)
actions = {agent_2: {Actions.FORWARD: 1}}
simple_jungle.step(actions)
simple_jungle.step(actions)
# They are now on the same cell
assert agent_1.position == agent_2.position
assert agent_1.angle == agent_2.angle
# We place a boulder in front of agents
simple_jungle.add_object(ElementsEnv.BOULDER, (5, 3))
# Agent 1 tries to go to the boulder but fails
actions = {agent_1: {Actions.FORWARD: 1}}
_, rew, done = simple_jungle.step(actions)
assert agent_1.position == agent_2.position
assert rew[agent_1] == Rewards.REWARD_COLLISION.value
# Now, agent 1 tries to climb first and then move forward
actions = {agent_1: {Actions.CLIMB: 1}}
simple_jungle.step(actions)
assert agent_1.on_shoulders
actions = {agent_1: {Actions.FORWARD: 1}}
simple_jungle.step(actions)
# Now, should be on same position as boulder
assert agent_1.position == (5, 3)
assert agent_1.on_shoulders is False
# After that, the agent can move forward on the next empty cell
actions = {agent_1: {Actions.FORWARD: 1}}
_, rew, _ = simple_jungle.step(actions)
assert agent_1.position == (5, 2)
assert rew[agent_1] == 0
def test_cut_tree():
# agent takes 3 actions and comes to a tree. once cut , the cell becomes empty
agent_1 = Agent(range_observation=4)
agent_2 = Agent(range_observation=6)
simple_jungle = EmptyJungle(size=11)
simple_jungle.add_agents(agent_1, agent_2, random_position=False)
simple_jungle.add_object(ElementsEnv.TREE, (5, 7))
simple_jungle.add_object(ElementsEnv.TREE, (3, 5))
actions = {
agent_1: {
Actions.FORWARD: 0,
Actions.ROTATE: 0
},
agent_2: {
Actions.FORWARD: 1,
Actions.ROTATE: 0
}
}
_, rew, done = simple_jungle.step(actions)
assert agent_2.wood_logs == 1
assert simple_jungle.cell_type(5, 7) == ElementsEnv.EMPTY.value
# assert increase in logs for agent 2 and tree converts to empty + agent has logs
actions = {
agent_1: {
Actions.FORWARD: 1,
Actions.ROTATE: -1
},
agent_2: {
Actions.FORWARD: 0,
Actions.ROTATE: 0
}
}
_, rew, done = simple_jungle.step(actions)
_, rew, done = simple_jungle.step(actions)
_, rew, done = simple_jungle.step(actions)
assert agent_1.wood_logs == 1
assert simple_jungle.cell_type(3, 5) == ElementsEnv.EMPTY.value
def test_agent_view():
# depending on the obstacle(tree) position , agents should be able to see / not see
# rivers, boulders , exits etc
agent_1 = Agent(range_observation=4)
agent_2 = Agent(range_observation=4)
simple_jungle = EmptyJungle(size=11)
simple_jungle.add_agents(agent_1, agent_2)
# directly left of agent 1
simple_jungle.add_object(ElementsEnv.TREE, (5, 3))
# should not be able to see this boulder
simple_jungle.add_object(ElementsEnv.BOULDER, (5, 2))
# should be able to see this river
simple_jungle.add_object(ElementsEnv.RIVER, (3, 3))
# directly right of agent 2
simple_jungle.add_object(ElementsEnv.TREE, (5, 8))
# should not be able to see this boulder
simple_jungle.add_object(ElementsEnv.BOULDER, (5, 9))
# should be able to see this river
simple_jungle.add_object(ElementsEnv.RIVER, (3, 9))
actions = {
agent_1: {
Actions.FORWARD: 0,
Actions.ROTATE: 0,
Actions.CLIMB: 0
},
agent_2: {
Actions.FORWARD: 0,
Actions.ROTATE: 0,
Actions.CLIMB: 0
}
}
obs, rew, done = simple_jungle.step(actions)
assert ElementsEnv.RIVER.value in obs[agent_1]
assert ElementsEnv.BOULDER.value not in obs[agent_1]
assert ElementsEnv.RIVER.value in obs[agent_2]
assert ElementsEnv.BOULDER.value not in obs[agent_2]
def test_initialization():
agent_1 = Agent()
agent_2 = Agent()
for size_envir in [11, 13, 15]:
simple_jungle = EmptyJungle(size=size_envir)
simple_jungle.add_agents(agent_1, agent_2)
# position should be in np coordinates (row, col)
possible_positions = (((size_envir - 1) / 2, (size_envir - 1) / 2 - 1),
((size_envir - 1) / 2, (size_envir - 1) / 2 + 1))
assert agent_1.position in possible_positions
assert agent_2.position in possible_positions
assert agent_1.position != agent_2.position
# angle is index of trigonometric angle (0, 1, ... to 5)
assert agent_1.angle in [3, 0]
assert agent_2.angle in [3, 0]
def test_obstacles_in_obs_diagonal():
agent_1 = Agent(range_observation=4)
agent_2 = Agent(range_observation=4)
simple_jungle = EmptyJungle(size=11)
simple_jungle.add_agents(agent_1, agent_2)
# bottom left diagonal for agent 1
simple_jungle.add_object(ElementsEnv.TREE, (6, 3))
# top right diagonal agent 2
simple_jungle.add_object(ElementsEnv.TREE, (3, 8))
actions = {}
obs, rew, done = simple_jungle.step(actions)
# @MG these assert statements were just to do some checks - they
# will not be part of the final codebase
assert agent_1.bottom_left_obstructed
assert agent_2.top_right_obstructed
def test_climb_action():
agent_1 = Agent(range_observation=4)
agent_2 = Agent(range_observation=4)
simple_jungle = EmptyJungle(size=11)
simple_jungle.add_agents(agent_1, agent_2)
# agents should do something and land on the same cell
# then they can climb
actions = {agent_1: {Actions.ROTATE: -1}}
simple_jungle.step(actions)
simple_jungle.step(actions)
simple_jungle.step(actions)
actions = {agent_1: {Actions.FORWARD: -1}}
simple_jungle.step(actions)
simple_jungle.step(actions)
# They are now on the same cell
assert agent_1.position == agent_2.position
actions = {agent_1: {Actions.CLIMB: 1}}
_, rew, done = simple_jungle.step(actions)
# black climbs, white does not move , observability increases, small neg reward for white as black is chubby. For
# now, lets say orientation is the direction white is facing ; so black, in climbing, changes its orientation to
# that of white
assert rew[agent_2] == Rewards.REWARD_CARRYING.value
assert agent_1.on_shoulders
# Finally, agent_2 moves, so agent_1 is not on shoulders anymore
# Agent 1 reeives a negative reward because it fell.
actions = {agent_2: {Actions.FORWARD: 1}}
_, rew, _ = simple_jungle.step(actions)
assert rew[agent_1] == Rewards.REWARD_FELL.value
assert not agent_1.on_shoulders
def test_approach_river_together():
agent_1 = Agent(range_observation=4)
agent_2 = Agent(range_observation=6)
simple_jungle = EmptyJungle(size=11)
simple_jungle.add_agents(agent_1, agent_2, random_position=False)
simple_jungle.add_object(ElementsEnv.RIVER, (4, 5))
agent_1.position, agent_1.angle = simple_jungle._starting_coordinates_1
agent_2.position, agent_2.angle = simple_jungle._starting_coordinates_2
actions = {
agent_1: {
Actions.FORWARD: 1,
Actions.ROTATE: 0
},
agent_2: {
Actions.FORWARD: 1,
Actions.ROTATE: 1
}
}
_, rew, done = simple_jungle.step(actions)
_, rew, done = simple_jungle.step(actions)
_, rew, done = simple_jungle.step(actions)
_, rew, done = simple_jungle.step(actions)
# both agents need to be at River, otherwise lone agent at river dies
assert rew[agent_2] == Rewards.REWARD_DROWN.value
# From MG test that agent 2 is dead -- eg maybe no value for agent 2 on grid
assert agent_2.done
assert not agent_1.done
def test_obs_new_jungle():
agent_1 = Agent(range_observation=4)
agent_2 = Agent(range_observation=4)
simple_jungle = EmptyJungle(size=11)
simple_jungle.add_agents(agent_1, agent_2)
# directly left of agent 1
simple_jungle.add_object(ElementsEnv.TREE, (5, 3))
# directly right of agent 2
simple_jungle.add_object(ElementsEnv.TREE, (5, 8))
actions = {}
obs, rew, done = simple_jungle.step(actions)
assert agent_1.left_view_obstructed
assert agent_2.right_view_obstructed
# directly below agent 1
simple_jungle.add_object(ElementsEnv.TREE, (8, 4))
# directly above agent 2
simple_jungle.add_object(ElementsEnv.TREE, (3, 7))
actions = {
agent_1: {
Actions.FORWARD: 1,
Actions.ROTATE: 1
},
agent_2: {
Actions.FORWARD: 1,
Actions.ROTATE: 1
}
}
obs, rew, done = simple_jungle.step(actions)
assert agent_1.bottom_view_obstructed
assert agent_2.top_view_obstructed
def test_obs():
# @MG I was just testing that observations are processed and returned
agent_1 = Agent(range_observation=4)
agent_2 = Agent(range_observation=4)
simple_jungle = EmptyJungle(size=11)
simple_jungle.add_agents(agent_1, agent_2)
actions = {
agent_1: {
Actions.FORWARD: 0,
Actions.ROTATE: -1,
Actions.CLIMB: 0
},
agent_2: {
Actions.FORWARD: 0,
Actions.ROTATE: 1,
Actions.CLIMB: 0
}
}
obs, rew, done = simple_jungle.step(actions)
def test__diagonal_view_bottom_right():
# test for top right
agent_1 = Agent(range_observation=4)
agent_2 = Agent(range_observation=4)
simple_jungle = EmptyJungle(size=11)
simple_jungle.add_agents(agent_1, agent_2)
# bottom left diagonal for agent 1
simple_jungle.add_object(ElementsEnv.TREE, (6, 7))
# should not be able to see around the diagonal
simple_jungle.add_object(ElementsEnv.RIVER, (6, 8))
simple_jungle.add_object(ElementsEnv.RIVER, (7, 7))
simple_jungle.add_object(ElementsEnv.RIVER, (7, 8))
# empty actions to pass to the step function
actions = {
agent_1: {
Actions.FORWARD: 0,
Actions.ROTATE: 0,
Actions.CLIMB: 0
},
agent_2: {
Actions.FORWARD: 0,
Actions.ROTATE: 0,
Actions.CLIMB: 0
}
}
obs, rew, done = simple_jungle.step(actions)
print(obs[agent_2])
assert agent_2.bottom_right_obstructed
assert ElementsEnv.RIVER.value not in obs[agent_2]
def test_movements():
agent_1 = Agent()
agent_2 = Agent()
simple_jungle = EmptyJungle(size=11)
simple_jungle.add_agents(agent_1, agent_2)
# First rotation, then forward, but the order in the actions dict doesn't matter.
actions = {
agent_1: {
Actions.FORWARD: 1,
Actions.ROTATE: 1
},
agent_2: {
Actions.FORWARD: 1,
Actions.ROTATE: -1
}
}
# Check positions on grid
pos_1 = agent_1.position
pos_2 = agent_2.position
angle_1 = agent_1.angle
angle_2 = agent_2.angle
# Perform the same action 6 times
# and it should go back to original position
# Basically, it is doing a circle
for _ in range(6):
simple_jungle.step(actions)
assert agent_1.position == pos_1
assert agent_1.angle == angle_1
assert agent_2.position == pos_2
assert agent_2.angle == angle_2
def test_collisions_with_obstacles():
# Agent 1 moves and collides with obstacles.
# Looks something like:
# . X .
# . . .
# . X A .
agent_1 = Agent(range_observation=4)
agent_2 = Agent(range_observation=6)
simple_jungle = EmptyJungle(size=11)
simple_jungle.add_agents(agent_1, agent_2, random_position=False)
simple_jungle.add_object(ElementsEnv.OBSTACLE, (5, 3))
simple_jungle.add_object(ElementsEnv.OBSTACLE, (3, 3))
# agent_1 moves forward, towards the object.
actions = {agent_1: {Actions.FORWARD: 1}}
_, rew, _ = simple_jungle.step(actions)
# agent keeps position and angle
assert agent_1.position == (5, 4)
assert agent_1.angle == 3
# agent receives reward for collision
assert rew[agent_1] == Rewards.REWARD_COLLISION.value
# agent_1 now rotates, then moves forward towards another object.
actions = {
agent_1: {
Actions.FORWARD: 1,
Actions.ROTATE: -1,
Actions.CLIMB: 0
},
agent_2: {
Actions.FORWARD: 0,
Actions.ROTATE: 0,
Actions.CLIMB: 0
}
}
_, rew, _ = simple_jungle.step(actions)
# first movement occurs without collision
assert agent_1.position == (4, 4)
assert rew[agent_1] == 0.0
assert agent_1.angle == 2
# agent_1 moves forward, now it should bump
actions = {
agent_1: {
Actions.FORWARD: 1,
Actions.ROTATE: 0,
Actions.CLIMB: 0
},
agent_2: {
Actions.FORWARD: 0,
Actions.ROTATE: 0,
Actions.CLIMB: 0
}
}
_, rew, _ = simple_jungle.step(actions)
assert agent_1.position == (4, 4)
assert rew[agent_1] == Rewards.REWARD_COLLISION.value
assert agent_1.angle == 2
def test_obs_cooperation_sequence():
# agent 2's view is obstructed
agent_1 = Agent(range_observation=4)
agent_2 = Agent(range_observation=4)
simple_jungle = EmptyJungle(size=11)
simple_jungle.add_agents(agent_1, agent_2)
print(agent_1.position, agent_1.angle)
print(agent_2.position, agent_2.angle)
simple_jungle.add_object(ElementsEnv.TREE, (5, 3))
actions = {}
obs, rew, done = simple_jungle.step(actions)
assert agent_1.left_view_obstructed
simple_jungle.add_object(ElementsEnv.TREE, (4, 7))
# should not be able to see around the diagonal
simple_jungle.add_object(ElementsEnv.RIVER, (4, 8))
simple_jungle.add_object(ElementsEnv.RIVER, (3, 7))
simple_jungle.add_object(ElementsEnv.RIVER, (3, 8))
# empty actions to pass to the step function
# agent 1 turns towards agent 2
actions = {
agent_1: {
Actions.FORWARD: 0,
Actions.ROTATE: -1,
Actions.CLIMB: 0
},
agent_2: {
Actions.FORWARD: 0,
Actions.ROTATE: 0,
Actions.CLIMB: 0
}
}
obs, rew, done = simple_jungle.step(actions)
obs, rew, done = simple_jungle.step(actions)
obs, rew, done = simple_jungle.step(actions)
assert agent_2.top_right_obstructed
assert ElementsEnv.RIVER.value not in obs[agent_2]
# agent 1 moves towards agent 2
actions = {
agent_1: {
Actions.FORWARD: 1,
Actions.ROTATE: 0,
Actions.CLIMB: 0
},
agent_2: {
Actions.FORWARD: 0,
Actions.ROTATE: 0,
Actions.CLIMB: 0
}
}
obs, rew, done = simple_jungle.step(actions)
obs, rew, done = simple_jungle.step(actions)
assert agent_1.position == (5, 6)
assert agent_2.position == (5, 6)
# agent 2 climbs on the shoulders of agent 1
actions = {
agent_1: {
Actions.FORWARD: 0,
Actions.ROTATE: 0,
Actions.CLIMB: 0
},
agent_2: {
Actions.FORWARD: 0,
Actions.ROTATE: 0,
Actions.CLIMB: 1
}
}
obs, rew, done = simple_jungle.step(actions)
assert agent_2.on_shoulders
# agent 2's observability is restored to full
assert ElementsEnv.RIVER.value in obs[agent_2]
def test_exits():
agent_1 = Agent(range_observation=4)
agent_2 = Agent(range_observation=6)
assert agent_1.done is False
assert agent_2.done is False
# we put exit towards an agent and move through it
actions = {
agent_1: {
Actions.FORWARD: 1,
Actions.ROTATE: 0
},
agent_2: {
Actions.FORWARD: 0,
Actions.ROTATE: 0
}
}
# exits provide 4 different rewards: LOW, AVERAGE, HIGH, VERY_HIGH
# There are 4 different kinds of exits:
# - EXIT_EASY: provide average reward
# - EXIT_DIFFICULT: provide high reward
# - EXIT_WHITE: provide very high reward to white, low reward to black
# - EXIT_BLACK: provide very high reward to black, low reward to white
# agent 1 takes easy exit.
simple_jungle = EmptyJungle(size=11)
simple_jungle.add_agents(agent_1, agent_2, random_position=False)
simple_jungle.add_object(ElementsEnv.EXIT_EASY, (5, 3))
_, rew, done = simple_jungle.step(actions)
assert rew[agent_1] == Rewards.REWARD_EXIT_AVERAGE.value
# agent 1 takes hard exit.
simple_jungle = EmptyJungle(size=11)
simple_jungle.add_agents(agent_1, agent_2, random_position=False)
simple_jungle.add_object(ElementsEnv.EXIT_DIFFICULT, (5, 3))
# when entering a new environment, agents done is reset.
assert agent_1.done is False
assert agent_2.done is False
_, rew, done = simple_jungle.step(actions)
assert rew[agent_1] == Rewards.REWARD_EXIT_HIGH.value
# If an agent takes the exit of its color, it receives a very high reward
actions = {
agent_1: {
Actions.FORWARD: 1,
Actions.ROTATE: 0
},
agent_2: {
Actions.FORWARD: 1,
Actions.ROTATE: 0
}
}
simple_jungle = EmptyJungle(size=11)
simple_jungle.add_agents(agent_1, agent_2, random_position=False)
if agent_1.color is WHITE:
simple_jungle.add_object(ElementsEnv.EXIT_WHITE, (5, 3))
simple_jungle.add_object(ElementsEnv.EXIT_BLACK, (5, 7))
else:
simple_jungle.add_object(ElementsEnv.EXIT_BLACK, (5, 3))
simple_jungle.add_object(ElementsEnv.EXIT_WHITE, (5, 7))
_, rew, done = simple_jungle.step(actions)
assert rew[agent_1] == Rewards.REWARD_EXIT_VERY_HIGH.value
assert rew[agent_1] == Rewards.REWARD_EXIT_VERY_HIGH.value
assert agent_1.done
assert agent_2.done
assert done
# If an agent takes the exit of the opposite color, it receives a low reward
simple_jungle = EmptyJungle(size=11)
simple_jungle.add_agents(agent_1, agent_2, random_position=False)
if agent_1.color is BLACK:
simple_jungle.add_object(ElementsEnv.EXIT_WHITE, (5, 3))
simple_jungle.add_object(ElementsEnv.EXIT_BLACK, (5, 7))
else:
simple_jungle.add_object(ElementsEnv.EXIT_BLACK, (5, 3))
simple_jungle.add_object(ElementsEnv.EXIT_WHITE, (5, 7))
_, rew, done = simple_jungle.step(actions)
assert rew[agent_1] == Rewards.REWARD_EXIT_LOW.value
assert rew[agent_1] == Rewards.REWARD_EXIT_LOW.value
def test_gameplay_exit():
# Game continues when one agent exits.
# Game terminates when both agents exit.
agent_1 = Agent(range_observation=4)
agent_2 = Agent(range_observation=6)
# agent 1 takes easy exit.
simple_jungle = EmptyJungle(size=11)
simple_jungle.add_agents(agent_1, agent_2, random_position=False)
simple_jungle.add_object(ElementsEnv.EXIT_EASY, (5, 3))
actions = {
agent_1: {
Actions.FORWARD: 1,
Actions.ROTATE: 0
},
agent_2: {
Actions.FORWARD: 0,
Actions.ROTATE: 0
}
}
_, rew, done = simple_jungle.step(actions)
assert done[agent_1] is True
assert done[agent_2] is False
assert agent_1.done is True
assert not agent_2.done
# agent 2 rotates then goes towards exit.
actions = {
agent_1: {
Actions.FORWARD: 0,
Actions.ROTATE: 0
},
agent_2: {
Actions.FORWARD: 0,
Actions.ROTATE: 1
}
}
simple_jungle.step(actions)
simple_jungle.step(actions)
simple_jungle.step(actions)
assert agent_1.done
assert not agent_2.done
actions = {
agent_1: {
Actions.FORWARD: 0,
Actions.ROTATE: 0
},
agent_2: {
Actions.FORWARD: 1,
Actions.ROTATE: 0
}
}
simple_jungle.add_object(ElementsEnv.EXIT_EASY, (5, 4))
simple_jungle.step(actions)
_, _, done = simple_jungle.step(actions)
assert agent_1.done
assert agent_2.done
assert done
def test_collision_with_tree():
# Agent 1 moves and collides with tree.
# Looks something like:
# . . . . . .
# . T T T . .
# . . . . A .
agent_1 = Agent()
agent_2 = Agent()
simple_jungle = EmptyJungle(size=11)
simple_jungle.add_agents(agent_1, agent_2, random_position=False)
simple_jungle.add_object(ElementsEnv.TREE, (4, 4))
simple_jungle.add_object(ElementsEnv.TREE, (4, 3))
simple_jungle.add_object(ElementsEnv.TREE, (4, 2))
# move to first tree
# agent_1 forward +1 and rotates -1
actions = {
agent_1: {
Actions.FORWARD: 1,
Actions.ROTATE: -1,
Actions.CLIMB: 0
},
agent_2: {
Actions.FORWARD: 0,
Actions.ROTATE: 0,
Actions.CLIMB: 0
}
}
obs, rew, done = simple_jungle.step(actions)
assert agent_1.position == (4, 4)
assert agent_1.angle == 2
assert rew[agent_1] == Rewards.REWARD_CUT_TREE.value
assert agent_1.wood_logs == 1
# move to second tree
# agent_1 forward +1 and rotates 1
actions = {
agent_1: {
Actions.FORWARD: 1,
Actions.ROTATE: 1,
Actions.CLIMB: 0
},
agent_2: {
Actions.FORWARD: 0,
Actions.ROTATE: 0,
Actions.CLIMB: 0
}
}
obs, rew, done = simple_jungle.step(actions)
assert agent_1.position == (4, 3)
assert agent_1.angle == 3
assert rew[agent_1] == Rewards.REWARD_CUT_TREE.value
assert agent_1.wood_logs == 2
# move to third tree
# agent_1 forward +1
actions = {
agent_1: {
Actions.FORWARD: 1,
Actions.ROTATE: 0,
Actions.CLIMB: 0
},
agent_2: {
Actions.FORWARD: 0,
Actions.ROTATE: 0,
Actions.CLIMB: 0
}
}
obs, rew, done = simple_jungle.step(actions)
assert agent_1.position == (4, 2)
assert agent_1.angle == 3
assert rew[agent_1] == Rewards.REWARD_CUT_TREE.value
assert agent_1.wood_logs == 2
from jungle.jungles.basic import FullJungle, RiverOnlyJungle , WhiteJungle, DifficultJungle
from jungle.jungles.rl import EasyExit
from jungle.agent import Agent
from jungle.utils import Actions
import matplotlib.pyplot as plt
import time
agent_1 = Agent(range_observation=5)
agent_2 = Agent(range_observation=5)
env = DifficultJungle(size=15)
env.add_agents(agent_1, agent_2)
t0 = time.time()
n_steps = 10000
for ts in range(n_steps):
# print(ts)
# print(agent_1.position, agent_2.position)
actions = {
agent_1: agent_1.get_random_actions(),
agent_2: agent_2.get_random_actions()
}
obs, rew, dones = env.step(actions)
print(n_steps / (time.time() - t0))
print(obs)
print(rew)