def test_tfa_runtime():
params = ParameterServer(
filename="tests/data/deterministic_scenario_test.json")
scenario_generation = DeterministicScenarioGeneration(num_scenarios=3,
random_seed=0,
params=params)
state_observer = ClosestAgentsObserver(params=params)
action_wrapper = DynamicModel(params=params)
evaluator = GoalReached(params=params)
viewer = MPViewer(params=params,
x_range=[-30, 30],
y_range=[-20, 40],
follow_agent_id=True) # use_world_bounds=True
runtimerl = RuntimeRL(action_wrapper=action_wrapper,
observer=state_observer,
evaluator=evaluator,
step_time=0.05,
viewer=viewer,
scenario_generator=scenario_generation)
tfa_env = TFAWrapper(runtimerl)
_ = tfa_env.reset()
utils.validate_py_environment(tfa_env, episodes=5)
_ = tf_py_environment.TFPyEnvironment(tfa_env)
def test_motion_primitives_concat_state(self):
params = ParameterServer(
filename="tests/data/deterministic_scenario_test.json")
scenario_generation = DeterministicScenarioGeneration(num_scenarios=3,
random_seed=0,
params=params)
state_observer = SimpleObserver(params=params)
action_wrapper = MotionPrimitives(params=params)
evaluator = GoalReached(params=params)
viewer = MPViewer(params=params,
x_range=[-30, 30],
y_range=[-40, 40],
use_world_bounds=True)
runtimerl = RuntimeRL(action_wrapper=action_wrapper,
observer=state_observer,
evaluator=evaluator,
step_time=0.2,
viewer=viewer,
scenario_generator=scenario_generation,
render=False)
for _ in range(0, 3):
runtimerl.reset()
for _ in range(0, 50): # run each scenario for 10 steps
action = action_wrapper.action_space.sample()
next_observed_state, reward, done, info = \
runtimerl.step(action)
if done:
print("State: {} \n Reward: {} \n Done {}, Info: {} \n \
=================================================" . \
format(next_observed_state, reward, done, info))
break
def test_runner():
params = ParameterServer(
filename="tests/data/deterministic_scenario_test.json")
base_dir = os.path.dirname(os.path.dirname(__file__))
params["BaseDir"] = base_dir
scenario_generation = DeterministicScenarioGeneration(num_scenarios=3,
random_seed=0,
params=params)
state_observer = ClosestAgentsObserver(params=params)
action_wrapper = DynamicModel(params=params)
evaluator = GoalReached(params=params)
viewer = MPViewer(params=params,
x_range=[-30, 30],
y_range=[-20, 40],
follow_agent_id=True)
runtimerl = RuntimeRL(action_wrapper=action_wrapper,
observer=state_observer,
evaluator=evaluator,
step_time=0.2,
viewer=viewer,
scenario_generator=scenario_generation,
render=False)
tfa_env = tf_py_environment.TFPyEnvironment(TFAWrapper(runtimerl))
sac_agent = SACAgent(tfa_env, params=params)
tfa_runner = TFARunner(tfa_env,
sac_agent,
params=params,
unwrapped_runtime=runtimerl)
tfa_runner.collect_initial_episodes()
# main functionalities
tfa_runner.train()
tfa_runner.visualize()
tfa_runner.evaluate()
def test_agent(self):
params = ParameterServer(
filename="tests/data/deterministic_scenario_test.json")
base_dir = os.path.dirname(os.path.dirname(__file__))
params["BaseDir"] = base_dir
scenario_generation = DeterministicScenarioGeneration(num_scenarios=2,
random_seed=0,
params=params)
state_observer = ClosestAgentsObserver(params=params)
action_wrapper = DynamicModel(params=params)
evaluator = GoalReached(params=params)
viewer = MPViewer(params=params,
x_range=[-30, 30],
y_range=[-20, 40],
follow_agent_id=True) # use_world_bounds=True
runtimerl = RuntimeRL(action_wrapper=action_wrapper,
observer=state_observer,
evaluator=evaluator,
step_time=0.05,
viewer=viewer,
scenario_generator=scenario_generation)
tfa_env = tf_py_environment.TFPyEnvironment(TFAWrapper(runtimerl))
sac_agent = SACAgent(tfa_env, params=params)
self.assertEqual(sac_agent._agent.name, "sac_agent")
sac_agent.reset()
# TODO(@hart): does not work because of read-only file-system
# sac_agent.save()
sac_agent.load()
def test_runtime_rl(self):
params = ParameterServer(
filename="tests/data/deterministic_scenario_test.json")
scenario_generation = DeterministicScenarioGeneration(num_scenarios=2,
random_seed=0,
params=params)
state_observer = SimpleObserver(params=params)
action_wrapper = DynamicModel(params=params)
evaluator = GoalReached(params=params)
viewer = MPViewer(params=params,
x_range=[-30, 30],
y_range=[-40, 40],
use_world_bounds=True) #use_world_bounds=True) #
runtimerl = RuntimeRL(action_wrapper=action_wrapper,
observer=state_observer,
evaluator=evaluator,
step_time=0.2,
viewer=viewer,
scenario_generator=scenario_generation,
render=False)
start_time = time.time()
for _ in range(0, 100):
runtimerl.reset()
done = False
reward = 0.
for _ in range(0, 50): # run each scenario for 10 steps
action = action_wrapper.action_space.sample(
) / 100 # to go straight
next_observed_state, reward, done, info = \
runtimerl.step(action)
# observer
self.assertEqual(len(next_observed_state), 16)
np.testing.assert_array_equal(
next_observed_state[0:4],
state_observer._normalize(
runtimerl._world.agents[100].state)[1:5])
np.testing.assert_array_equal(
next_observed_state[4:8],
state_observer._normalize(
runtimerl._world.agents[101].state)[1:5])
if done:
print("State: {} \n Reward: {} \n Done {}, Info: {} \n \
=================================================" . \
format(next_observed_state, reward, done, info))
break
# must assert to equal as the agent reaches the goal in the
# specified number of steps
self.assertEqual(done, True)
# goal must have been reached which returns a reward of 1.
self.assertEqual(reward, 1.)
self.assertEqual(runtimerl._world.agents[100].id, 100)
self.assertEqual(runtimerl._world.agents[101].id, 101)
end_time = time.time()
print("100 runs took {}s.".format(str(end_time - start_time)))
def __init__(self, params=ParameterServer(), eval_agent=None):
GoalReached.__init__(self, params, eval_agent)
self._next_goal_definition = -1
self._last_distance = None
self._goal_number = 0
def __init__(self,
params=ParameterServer(),
eval_agent=None):
GoalReached.__init__(self,
params,
eval_agent)
def __init__(self, params=ParameterServer(), eval_agent=None):
GoalReached.__init__(self, params, eval_agent)
self._last_goal_id = -1
self._reached_goal_in_last_step = False
# TODO: use from config file
self._intermediate_goal_reward = 1.0