diff --git a/examples/simple_example_3.py b/examples/simple_example_3.py
index 1661ef65a9a33f3b44a098caaf83317919722398..e015b3c88cf05a8d047f15dfaf88e8a2fd9ce789 100644
--- a/examples/simple_example_3.py
+++ b/examples/simple_example_3.py
@@ -2,7 +2,7 @@ import random
import numpy as np
-from flatland.envs.generators import random_rail_generator
+from flatland.envs.generators import random_rail_generator, complex_rail_generator
from flatland.envs.observations import TreeObsForRailEnv
from flatland.envs.rail_env import RailEnv
from flatland.utils.rendertools import RenderTool
diff --git a/flatland/envs/agent_utils.py b/flatland/envs/agent_utils.py
index 8e9ffb99d06176416dfbe2b65bcca09723b1a56c..aa46aecd4b69b6a13b11b63223123b16dd69e3ac 100644
--- a/flatland/envs/agent_utils.py
+++ b/flatland/envs/agent_utils.py
@@ -28,19 +28,34 @@ class EnvAgentStatic(object):
position = attrib()
direction = attrib()
target = attrib()
- moving = attrib()
-
- def __init__(self, position, direction, target, moving=False):
+ moving = attrib(default=False)
+ # speed_data: speed is added to position_fraction on each moving step, until position_fraction>=1.0,
+ # after which 'transition_action_on_cellexit' is executed (equivalent to executing that action in the previous
+ # cell if speed=1, as default)
+ speed_data = attrib(default=dict({'position_fraction': 0.0, 'speed': 1.0, 'transition_action_on_cellexit': 0}))
+
+ def __init__(self,
+ position,
+ direction,
+ target,
+ moving=False,
+ speed_data={'position_fraction': 0.0, 'speed': 1.0, 'transition_action_on_cellexit': 0}):
self.position = position
self.direction = direction
self.target = target
self.moving = moving
+ self.speed_data = speed_data
@classmethod
- def from_lists(cls, positions, directions, targets):
+ def from_lists(cls, positions, directions, targets, speeds=None):
""" Create a list of EnvAgentStatics from lists of positions, directions and targets
"""
- return list(starmap(EnvAgentStatic, zip(positions, directions, targets, [False] * len(positions))))
+ speed_datas = []
+ for i in range(len(positions)):
+ speed_datas.append({'position_fraction': 0.0,
+ 'speed': speeds[i] if speeds is not None else 1.0,
+ 'transition_action_on_cellexit': 0})
+ return list(starmap(EnvAgentStatic, zip(positions, directions, targets, [False] * len(positions), speed_datas)))
def to_list(self):
@@ -54,7 +69,7 @@ class EnvAgentStatic(object):
if type(lTarget) is np.ndarray:
lTarget = lTarget.tolist()
- return [lPos, int(self.direction), lTarget, int(self.moving)]
+ return [lPos, int(self.direction), lTarget, int(self.moving), self.speed_data]
@attrs
@@ -78,7 +93,7 @@ class EnvAgent(EnvAgentStatic):
def to_list(self):
return [
self.position, self.direction, self.target, self.handle,
- self.old_direction, self.old_position, self.moving]
+ self.old_direction, self.old_position, self.moving, self.speed_data]
@classmethod
def from_static(cls, oStatic):
diff --git a/flatland/envs/generators.py b/flatland/envs/generators.py
index f644bc120d4b514f1c54e0330cfc8dc4654a4f4e..ca14667424d2c93d1466e3b7e96c2e5c1fbd41e5 100644
--- a/flatland/envs/generators.py
+++ b/flatland/envs/generators.py
@@ -18,7 +18,7 @@ def empty_rail_generator():
rail_array = grid_map.grid
rail_array.fill(0)
- return grid_map, [], [], []
+ return grid_map, [], [], [], []
return generator
@@ -75,8 +75,9 @@ def complex_rail_generator(nr_start_goal=1, nr_extra=100, min_dist=20, max_dist=
while nr_created < nr_start_goal and created_sanity < sanity_max:
all_ok = False
for _ in range(sanity_max):
- start = (np.random.randint(0, width), np.random.randint(0, height))
- goal = (np.random.randint(0, height), np.random.randint(0, height))
+ start = (np.random.randint(0, height), np.random.randint(0, width))
+ goal = (np.random.randint(0, height), np.random.randint(0, width))
+
# check to make sure start,goal pos is empty?
if rail_array[goal] != 0 or rail_array[start] != 0:
continue
@@ -121,8 +122,8 @@ def complex_rail_generator(nr_start_goal=1, nr_extra=100, min_dist=20, max_dist=
while nr_created < nr_extra and created_sanity < sanity_max:
all_ok = False
for _ in range(sanity_max):
- start = (np.random.randint(0, width), np.random.randint(0, height))
- goal = (np.random.randint(0, height), np.random.randint(0, height))
+ start = (np.random.randint(0, height), np.random.randint(0, width))
+ goal = (np.random.randint(0, height), np.random.randint(0, width))
# check to make sure start,goal pos are not empty
if rail_array[goal] == 0 or rail_array[start] == 0:
continue
@@ -139,7 +140,7 @@ def complex_rail_generator(nr_start_goal=1, nr_extra=100, min_dist=20, max_dist=
agents_target = [sg[1] for sg in start_goal[:num_agents]]
agents_direction = start_dir[:num_agents]
- return grid_map, agents_position, agents_direction, agents_target
+ return grid_map, agents_position, agents_direction, agents_target, [1.0]*len(agents_position)
return generator
@@ -183,7 +184,7 @@ def rail_from_manual_specifications_generator(rail_spec):
rail,
num_agents)
- return rail, agents_position, agents_direction, agents_target
+ return rail, agents_position, agents_direction, agents_target, [1.0]*len(agents_position)
return generator
@@ -209,7 +210,7 @@ def rail_from_GridTransitionMap_generator(rail_map):
rail_map,
num_agents)
- return rail_map, agents_position, agents_direction, agents_target
+ return rail_map, agents_position, agents_direction, agents_target, [1.0]*len(agents_position)
return generator
@@ -482,6 +483,6 @@ def random_rail_generator(cell_type_relative_proportion=[1.0] * 11):
return_rail,
num_agents)
- return return_rail, agents_position, agents_direction, agents_target
+ return return_rail, agents_position, agents_direction, agents_target, [1.0]*len(agents_position)
return generator
diff --git a/flatland/envs/rail_env.py b/flatland/envs/rail_env.py
index c22e1c5120b54a170f9c59bb54c7666ca910f086..8cf6d52f383ec8f4e271eb0765d32bc0c763307a 100644
--- a/flatland/envs/rail_env.py
+++ b/flatland/envs/rail_env.py
@@ -73,7 +73,7 @@ class RailEnv(Environment):
random_rail_generator : generate a random rail of given size
rail_from_GridTransitionMap_generator(rail_map) : generate a rail from
a GridTransitionMap object
- rail_from_manual_specifications_generator(rail_spec) : generate a rail from
+ rail_from_manual_sp ecifications_generator(rail_spec) : generate a rail from
a rail specifications array
TODO: generate_rail_from_saved_list or from list of ndarray bitmaps ---
width : int
@@ -101,7 +101,6 @@ class RailEnv(Environment):
self.action_space = [1]
self.observation_space = self.obs_builder.observation_space # updated on resets?
- self.actions = [0] * number_of_agents
self.rewards = [0] * number_of_agents
self.done = False
@@ -152,7 +151,7 @@ class RailEnv(Environment):
self.rail = tRailAgents[0]
if replace_agents:
- self.agents_static = EnvAgentStatic.from_lists(*tRailAgents[1:4])
+ self.agents_static = EnvAgentStatic.from_lists(*tRailAgents[1:5])
self.restart_agents()
@@ -193,28 +192,26 @@ class RailEnv(Environment):
for iAgent in range(self.get_num_agents()):
agent = self.agents[iAgent]
- if iAgent not in action_dict: # no action has been supplied for this agent
- if agent.moving:
- # Keep moving
- # Change MOVE_FORWARD to DO_NOTHING
- action_dict[iAgent] = RailEnvActions.DO_NOTHING
- else:
- action_dict[iAgent] = RailEnvActions.DO_NOTHING
-
if self.dones[iAgent]: # this agent has already completed...
continue
- action = action_dict[iAgent]
- if action < 0 or action > len(RailEnvActions):
- print('ERROR: illegal action=', action,
- 'for agent with index=', iAgent)
- return
+ if iAgent not in action_dict: # no action has been supplied for this agent
+ action_dict[iAgent] = RailEnvActions.DO_NOTHING
+
+ if action_dict[iAgent] < 0 or action_dict[iAgent] > len(RailEnvActions):
+ print('ERROR: illegal action=', action_dict[iAgent],
+ 'for agent with index=', iAgent,
+ '"DO NOTHING" will be executed instead')
+ action_dict[iAgent] = RailEnvActions.DO_NOTHING
+
+ action = action_dict[iAgent]
if action == RailEnvActions.DO_NOTHING and agent.moving:
# Keep moving
action = RailEnvActions.MOVE_FORWARD
- if action == RailEnvActions.STOP_MOVING and agent.moving:
+ if action == RailEnvActions.STOP_MOVING and agent.moving and agent.speed_data['position_fraction'] < 0.01:
+ # Only allow halting an agent on entering new cells.
agent.moving = False
self.rewards_dict[iAgent] += stop_penalty
@@ -223,47 +220,78 @@ class RailEnv(Environment):
agent.moving = True
self.rewards_dict[iAgent] += start_penalty
- if action != RailEnvActions.DO_NOTHING and action != RailEnvActions.STOP_MOVING:
- cell_isFree, new_cell_isValid, new_direction, new_position, transition_isValid = \
- self._check_action_on_agent(action, agent)
- if all([new_cell_isValid, transition_isValid, cell_isFree]):
- agent.old_direction = agent.direction
- agent.old_position = agent.position
- agent.position = new_position
- agent.direction = new_direction
- else:
- # Logic: if the chosen action is invalid,
- # and it was LEFT or RIGHT, and the agent was moving, then keep moving FORWARD.
- if (action == RailEnvActions.MOVE_LEFT or action == RailEnvActions.MOVE_RIGHT) and agent.moving:
- cell_isFree, new_cell_isValid, new_direction, new_position, transition_isValid = \
- self._check_action_on_agent(RailEnvActions.MOVE_FORWARD, agent)
-
- if all([new_cell_isValid, transition_isValid, cell_isFree]):
- agent.old_direction = agent.direction
- agent.old_position = agent.position
- agent.position = new_position
- agent.direction = new_direction
+ # Now perform a movement.
+ # If the agent is in an initial position within a new cell (agent.speed_data['position_fraction']<eps)
+ # store the desired action in `transition_action_on_cellexit' (only if the desired transition is
+ # allowed! otherwise DO_NOTHING!)
+ # Then in any case (if agent.moving) and the `transition_action_on_cellexit' is valid, increment the
+ # position_fraction by the speed of the agent (regardless of action taken, as long as no
+ # STOP_MOVING, but that makes agent.moving=False)
+ # If the new position fraction is >= 1, reset to 0, and perform the stored
+ # transition_action_on_cellexit
+
+ # If the agent can make an action
+ action_selected = False
+ if agent.speed_data['position_fraction'] < 0.01:
+ if action != RailEnvActions.DO_NOTHING and action != RailEnvActions.STOP_MOVING:
+ cell_isFree, new_cell_isValid, new_direction, new_position, transition_isValid = \
+ self._check_action_on_agent(action, agent)
+
+ if all([new_cell_isValid, transition_isValid, cell_isFree]):
+ agent.speed_data['transition_action_on_cellexit'] = action
+ action_selected = True
+
+ else:
+ # But, if the chosen invalid action was LEFT/RIGHT, and the agent is moving,
+ # try to keep moving forward!
+ if (action == RailEnvActions.MOVE_LEFT or action == RailEnvActions.MOVE_RIGHT) and agent.moving:
+ cell_isFree, new_cell_isValid, new_direction, new_position, transition_isValid = \
+ self._check_action_on_agent(RailEnvActions.MOVE_FORWARD, agent)
+
+ if all([new_cell_isValid, transition_isValid, cell_isFree]):
+ agent.speed_data['transition_action_on_cellexit'] = RailEnvActions.MOVE_FORWARD
+ action_selected = True
+
+ else:
+ # TODO: an invalid action was chosen after entering the cell. The agent cannot move.
+ self.rewards_dict[iAgent] += invalid_action_penalty
+ agent.moving = False
+ self.rewards_dict[iAgent] += stop_penalty
+ continue
else:
- # the action was not valid, add penalty
+ # TODO: an invalid action was chosen after entering the cell. The agent cannot move.
self.rewards_dict[iAgent] += invalid_action_penalty
+ agent.moving = False
+ self.rewards_dict[iAgent] += stop_penalty
+ continue
- else:
- # the action was not valid, add penalty
- self.rewards_dict[iAgent] += invalid_action_penalty
+ if agent.moving and (action_selected or agent.speed_data['position_fraction'] >= 0.01):
+ agent.speed_data['position_fraction'] += agent.speed_data['speed']
+
+ if agent.speed_data['position_fraction'] >= 1.0:
+ agent.speed_data['position_fraction'] = 0.0
+
+ # Perform stored action to transition to the next cell
+
+ # Now 'transition_action_on_cellexit' will be guaranteed to be valid; it was checked on entering
+ # the cell
+ cell_isFree, new_cell_isValid, new_direction, new_position, transition_isValid = \
+ self._check_action_on_agent(agent.speed_data['transition_action_on_cellexit'], agent)
+ agent.old_direction = agent.direction
+ agent.old_position = agent.position
+ agent.position = new_position
+ agent.direction = new_direction
if np.equal(agent.position, agent.target).all():
self.dones[iAgent] = True
else:
- self.rewards_dict[iAgent] += step_penalty
+ self.rewards_dict[iAgent] += step_penalty * agent.speed_data['speed']
# Check for end of episode + add global reward to all rewards!
if np.all([np.array_equal(agent2.position, agent2.target) for agent2 in self.agents]):
self.dones["__all__"] = True
self.rewards_dict = [0 * r + global_reward for r in self.rewards_dict]
- # Reset the step actions (in case some agent doesn't 'register_action'
- # on the next step)
- self.actions = [0] * self.get_num_agents()
return self._get_observations(), self.rewards_dict, self.dones, {}
def _check_action_on_agent(self, action, agent):
@@ -271,6 +299,7 @@ class RailEnv(Environment):
# cell used to check for invalid actions
new_direction, transition_isValid = self.check_action(agent, action)
new_position = get_new_position(agent.position, new_direction)
+
# Is it a legal move?
# 1) transition allows the new_direction in the cell,
# 2) the new cell is not empty (case 0),
@@ -281,11 +310,13 @@ class RailEnv(Environment):
np.clip(new_position, [0, 0], [self.height - 1, self.width - 1]))
and # check the new position has some transitions (ie is not an empty cell)
self.rail.get_transitions(new_position) > 0)
+
# If transition validity hasn't been checked yet.
if transition_isValid is None:
transition_isValid = self.rail.get_transition(
(*agent.position, agent.direction),
new_direction)
+
# Check the new position is not the same as any of the existing agent positions
# (including itself, for simplicity, since it is moving)
cell_isFree = not np.any(
diff --git a/tests/__init__.py b/tests/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/tests/test_environments.py b/tests/test_environments.py
index 11f0acba2fd54df63c62047f8559897e7d222e72..aa24467dd1d548a2b68a408f300089ee8135c639 100644
--- a/tests/test_environments.py
+++ b/tests/test_environments.py
@@ -3,7 +3,7 @@
import numpy as np
from flatland.core.transition_map import GridTransitionMap
-from flatland.core.transitions import Grid4Transitions
+from flatland.core.transitions import Grid4Transitions, RailEnvTransitions
from flatland.envs.agent_utils import EnvAgent
from flatland.envs.generators import complex_rail_generator
from flatland.envs.generators import rail_from_GridTransitionMap_generator
@@ -53,7 +53,7 @@ def test_rail_environment_single_agent():
# | | |
# \_/\_/
- transitions = Grid4Transitions([])
+ transitions = RailEnvTransitions()
vertical_line = cells[1]
south_symmetrical_switch = cells[6]
north_symmetrical_switch = transitions.rotate_transition(south_symmetrical_switch, 180)
@@ -107,6 +107,7 @@ def test_rail_environment_single_agent():
if prev_pos != pos:
valid_active_actions_done += 1
+
# After 6 movements on this railway network, the train should be back
# to its original height on the map.
assert (initial_pos[0] == agent.position[0])
@@ -121,9 +122,9 @@ def test_rail_environment_single_agent():
action = np.random.randint(4)
_, _, dones, _ = rail_env.step({0: action})
-
done = dones['__all__']
+test_rail_environment_single_agent()
def test_dead_end():
transitions = Grid4Transitions([])