def test_create_sac_agent(self):
params = ParameterServer()
map_interface = self.map_interface
observer = NearestAgentsObserver()
env = ExternalRuntime(map_interface=map_interface,
observer=observer,
params=params)
env.ml_behavior = BehaviorContinuousML(params)
sac_agent = BehaviorSACAgent(environment=env, params=params)
env.ml_behavior = sac_agent
self.assertTrue(isinstance(env.ml_behavior, BehaviorSACAgent))
def test_generate_ego_trajectory(self):
params = ParameterServer()
env = self.create_runtime_and_setup_empty_world(params)
sac_agent = BehaviorSACAgent(environment=env, params=params)
env.ml_behavior = sac_agent
# add ego
state = np.array([0, 0, 0, 0, 0, 0])
goal_line = Line2d(np.array([[0., 0.], [1., 1.]]))
env.addEgoAgent(state, goal_line)
N = 10
state_traj, action_traj = env.generateTrajectory(0.2, N)
env._viewer.drawTrajectory(state_traj)
env.render()
self.assertEqual(len(state_traj), N)
def generate_test_trajectory(self):
"""Generates a test trajectory from a SAC agent
"""
sac_agent = BehaviorSACAgent(environment=self.env,
params=self.params)
self.env.ml_behavior = sac_agent
runner = SACRunnerGenerator(params=self.params,
environment=self.env,
agent=sac_agent)
expert_trajectories = runner.GenerateExpertTrajectories(
num_trajectories=1, render=False)
dirname = "test_expert_trajectories"
Path(dirname).mkdir(parents=True, exist_ok=True)
filename = os.path.join(dirname, "test_expert_trajectory.jblb")
joblib.dump(expert_trajectories[0], filename)
return dirname
def test_configurable_blueprint(self):
params = ParameterServer(
filename="bark_ml/tests/data/highway_merge_configurable.json")
# continuous model
ml_behavior = BehaviorContinuousML(params=params)
bp = ConfigurableScenarioBlueprint(params=params,
ml_behavior=ml_behavior)
env = SingleAgentRuntime(blueprint=bp, render=False)
# agent
sac_agent = BehaviorSACAgent(environment=env, params=params)
env.ml_behavior = sac_agent
# test run
env.reset()
for _ in range(0, 5):
action = np.random.randint(low=0, high=3)
observed_next_state, reward, done, info = env.step(action)
def test_tracing_bark_world(self):
params = ParameterServer()
bp = ContinuousHighwayBlueprint(params)
tracer = Tracer()
env = SingleAgentRuntime(blueprint=bp, render=False)
sac_agent = BehaviorSACAgent(environment=env, params=params)
env.ml_behavior = sac_agent
# NOTE: this also tests if a BARK agent is self-contained
env.ml_behavior.set_actions_externally = False
env.reset()
bark_world = env._world
for j in range(0, 2):
for i in range(0, 5):
bark_world.Step(0.2)
eval_dict = bark_world.Evaluate()
tracer.Trace(eval_dict, num_episode=j)
self.assertEqual(len(tracer._states), 10)
def test_behavior_wrapping(self):
# create scenario
params = ParameterServer()
bp = ContinuousHighwayBlueprint(params,
num_scenarios=10,
random_seed=0)
env = SingleAgentRuntime(blueprint=bp, render=False)
ml_behaviors = []
ml_behaviors.append(BehaviorPPOAgent(environment=env, params=params))
ml_behaviors.append(BehaviorSACAgent(environment=env, params=params))
for ml_behavior in ml_behaviors:
# set agent
env.ml_behavior = ml_behavior
env.reset()
done = False
while done is False:
action = np.random.uniform(low=-0.1, high=0.1, size=(2, ))
observed_next_state, reward, done, info = env.step(action)
print(
f"Observed state: {observed_next_state}, Reward: {reward}, Done: {done}"
)
# action is set externally
ml_behavior._set_action_externally = True
agent_id = list(env._world.agents.keys())[0]
observed_world = env._world.Observe([agent_id])[0]
action = np.random.uniform(low=-0.1, high=0.1, size=(2, ))
ml_behavior.ActionToBehavior(action)
a = ml_behavior.Plan(0.2, observed_world)
action = np.random.uniform(low=-0.1, high=0.1, size=(2, ))
ml_behavior.ActionToBehavior(action)
b = ml_behavior.Plan(0.2, observed_world)
self.assertEqual(np.any(np.not_equal(a, b)), True)
# action will be calculated within the Plan(..) fct.
a = ml_behavior.Plan(0.2, observed_world)
b = ml_behavior.Plan(0.2, observed_world)
np.testing.assert_array_equal(a, b)
def test_behavior_wrapping(self):
# create scenario
params = ParameterServer()
bp = ContinuousHighwayBlueprint(params,
number_of_senarios=10,
random_seed=0)
env = SingleAgentRuntime(blueprint=bp, render=True)
ml_behaviors = []
ml_behaviors.append(BehaviorPPOAgent(environment=env, params=params))
ml_behaviors.append(BehaviorSACAgent(environment=env, params=params))
for ml_behavior in ml_behaviors:
# set agent
env.ml_behavior = ml_behavior
env.reset()
done = False
while done is False:
action = np.random.uniform(low=-0.1, high=0.1, size=(2, ))
observed_next_state, reward, done, info = env.step(action)
print(
f"Observed state: {observed_next_state}, Reward: {reward}, Done: {done}"
)
def test_sac_agent(self):
params = ParameterServer()
env = gym.make("highway-v0")
sac_agent = BehaviorSACAgent(environment=env, params=params)
ppo_agent = BehaviorPPOAgent(environment=env, params=params)
behaviors = [ppo_agent, sac_agent]
for ml_agent in behaviors:
env.ml_behavior = ml_agent
env.reset()
eval_id = env._scenario._eval_agent_ids[0]
cloned_world = env._world.Copy()
self.assertEqual(env._world.agents[eval_id].behavior_model,
ml_agent)
for _ in range(0, 5):
env._world.Step(0.2)
self.assertEqual(cloned_world.agents[eval_id].behavior_model,
ml_agent)
for _ in range(0, 5):
cloned_world.Step(0.2)
for cloned_agent, agent in zip(env._world.agents.values(),
cloned_world.agents.values()):
# NOTE: should be the same as mean is taken from the agents
np.testing.assert_array_equal(cloned_agent.state, agent.state)
def test_agent_wrapping(self):
params = ParameterServer()
env = gym.make("highway-v0")
env.reset()
agent = BehaviorPPOAgent(environment=env, params=params)
agent = BehaviorSACAgent(environment=env, params=params)