def get_agent(self, ego, policy, max_episode_steps):
observation_adapter = None
if ego:
config = get_agent_config_by_type(policy)
agent_spec = AgentSpec(
interface=config["interface"],
agent_params=dict(config["policy"], checkpoint_dir=ego_model),
agent_builder=config["policy_class"],
)
agent_spec.interface.max_episode_steps = max_episode_steps
observation_adapter = IntersectionAdapter(
agent_id="AGENT_007",
social_vehicle_config=config["social_vehicle_config"],
timestep_sec=config["env"]["timestep_sec"],
**config["other"],
)
else:
# Lane Following agent
agent_spec = AgentSpec(
interface=AgentInterface(
max_episode_steps=max_episode_steps, # 10 mins
waypoints=True,
action=ActionSpaceType.Lane,
debug=False,
neighborhood_vehicles=NeighborhoodVehicles(radius=2000),
),
agent_builder=DefaultPolicy,
)
return agent_spec, observation_adapter
def main(scenarios, sim_name, headless, num_episodes, seed, max_episode_steps=None):
agent_spec = AgentSpec(
interface=AgentInterface.from_type(
AgentType.LanerWithSpeed, max_episode_steps=max_episode_steps
),
agent_builder=ChaseViaPointsAgent,
)
env = gym.make(
"smarts.env:hiway-v0",
scenarios=scenarios,
agent_specs={AGENT_ID: agent_spec},
sim_name=sim_name,
headless=headless,
visdom=False,
timestep_sec=0.1,
sumo_headless=True,
seed=seed,
# zoo_addrs=[("10.193.241.236", 7432)], # Sample server address (ip, port), to distribute social agents in remote server.
# envision_record_data_replay_path="./data_replay",
)
for episode in episodes(n=num_episodes):
agent = agent_spec.build_agent()
observations = env.reset()
episode.record_scenario(env.scenario_log)
dones = {"__all__": False}
while not dones["__all__"]:
agent_obs = observations[AGENT_ID]
agent_action = agent.act(agent_obs)
observations, rewards, dones, infos = env.step({AGENT_ID: agent_action})
episode.record_step(observations, rewards, dones, infos)
env.close()
def main(scenarios, headless, seed):
scenarios_iterator = Scenario.scenario_variations(scenarios, [])
for _ in scenarios:
scenario = next(scenarios_iterator)
agent_missions = scenario.discover_missions_of_traffic_histories()
for agent_id, mission in agent_missions.items():
scenario.set_ego_missions({agent_id: mission})
agent_spec = AgentSpec(
interface=AgentInterface.from_type(AgentType.Laner,
max_episode_steps=None),
agent_builder=KeepLaneAgent,
)
agent = agent_spec.build_agent()
smarts = SMARTS(
agent_interfaces={agent_id: agent_spec.interface},
traffic_sim=SumoTrafficSimulation(headless=True,
auto_start=True),
envision=Envision(),
)
observations = smarts.reset(scenario)
dones = {agent_id: False}
while not dones[agent_id]:
agent_obs = observations[agent_id]
agent_action = agent.act(agent_obs)
observations, rewards, dones, infos = smarts.step(
{agent_id: agent_action})
def main(scenarios, headless, num_episodes, seed):
agent_spec = AgentSpec(
interface=AgentInterface.from_type(AgentType.Laner, max_episode_steps=None),
agent_builder=KeepLaneAgent,
)
env = gym.make(
"smarts.env:hiway-v0",
scenarios=scenarios,
agent_specs={AGENT_ID: agent_spec},
headless=headless,
visdom=False,
timestep_sec=0.1,
sumo_headless=True,
seed=seed,
# envision_record_data_replay_path="./data_replay",
)
for episode in episodes(n=num_episodes):
agent = agent_spec.build_agent()
observations = env.reset()
episode.record_scenario(env.scenario_log)
dones = {"__all__": False}
while not dones["__all__"]:
agent_obs = observations[AGENT_ID]
agent_action = agent.act(agent_obs)
observations, rewards, dones, infos = env.step({AGENT_ID: agent_action})
episode.record_step(observations, rewards, dones, infos)
env.close()
def main(
scenarios,
headless,
num_episodes,
seed,
):
agent_spec = AgentSpec(
interface=AgentInterface.from_type(
AgentType.StandardWithAbsoluteSteering, max_episode_steps=3000),
policy_builder=HumanKeyboardPolicy,
)
env = gym.make(
"smarts.env:hiway-v0",
scenarios=scenarios,
agent_specs={AGENT_ID: agent_spec},
headless=headless,
timestep_sec=0.1,
seed=seed,
)
for episode in episodes(n=num_episodes):
agent = agent_spec.build_agent()
observations = env.reset()
episode.record_scenario(env.scenario_log)
dones = {"__all__": False}
while not dones["__all__"]:
agent_obs = observations[AGENT_ID]
agent_action = agent.act(agent_obs)
observations, rewards, dones, infos = env.step(
{AGENT_ID: agent_action})
episode.record_step(observations, rewards, dones, infos)
env.close()
def test_graceful_interrupt(monkeypatch):
"""SMARTS should only throw a KeyboardInterript exception."""
agent_spec = AgentSpec(
interface=AgentInterface.from_type(AgentType.Laner),
agent_builder=lambda: Agent.from_function(lambda _: "keep_lane"),
)
agent = agent_spec.build_agent()
env = build_env(agent_spec)
with pytest.raises(KeyboardInterrupt):
obs = env.reset()
# To simulate a user interrupting the sim (e.g. ctrl-c). We just need to
# hook in to some function that SMARTS calls internally (like this one).
with mock.patch(
"smarts.core.sensors.Sensors.observe", side_effect=KeyboardInterrupt
):
for episode in range(10):
obs, _, _, _ = env.step({AGENT_ID: agent.act(obs)})
assert episode == 0, "SMARTS should have been interrupted, ending early"
with pytest.raises(SMARTSNotSetupError):
env.step({AGENT_ID: agent.act(obs)})
def test_building_agent_with_dict_params():
agent_spec = AgentSpec(
agent_params={"y": 2, "x": 1},
agent_builder=lambda x, y: Agent.from_function(lambda _: x / y),
)
agent = agent_spec.build_agent()
assert agent.act("dummy observation") == 1 / 2
def test_building_agent_with_tuple_params():
agent_spec = AgentSpec(
agent_params=(32, 41),
agent_builder=lambda x, y: Agent.from_function(lambda _: (x, y)),
)
agent = agent_spec.build_agent()
assert agent.act("dummy observation") == (32, 41)
def test_building_agent_with_list_or_tuple_params():
agent_spec = AgentSpec(
policy_params=[32, 41],
policy_builder=lambda x, y: AgentPolicy.from_function(lambda _:
(x, y)),
)
agent = agent_spec.build_agent()
assert agent.act("dummy observation") == (32, 41)
def train(training_scenarios, evaluation_scenarios, sim_name, headless,
num_episodes, seed):
agent_params = {"input_dims": 4, "hidden_dims": 7, "output_dims": 3}
agent_spec = AgentSpec(
interface=AgentInterface.from_type(AgentType.Standard,
max_episode_steps=5000),
agent_params=agent_params,
agent_builder=PyTorchAgent,
observation_adapter=observation_adapter,
)
env = gym.make(
"smarts.env:hiway-v0",
scenarios=training_scenarios,
agent_specs={AGENT_ID: agent_spec},
sim_name=sim_name,
headless=headless,
fixed_timestep_sec=0.1,
seed=seed,
)
steps = 0
for episode in episodes(n=num_episodes):
agent = agent_spec.build_agent()
observations = env.reset()
episode.record_scenario(env.scenario_log)
dones = {"__all__": False}
while not dones["__all__"]:
agent_obs = observations[AGENT_ID]
agent_action = agent.act(agent_obs)
observations, rewards, dones, infos = env.step(
{AGENT_ID: agent_action})
episode.record_step(observations, rewards, dones, infos)
steps += 1
if steps % 500 == 0:
print("Evaluating agent")
# We construct an evaluation agent based on the saved
# state of the agent in training.
model_path = tempfile.mktemp()
agent.save(model_path)
eval_agent_spec = agent_spec.replace(
agent_params=dict(agent_params, model_path=model_path))
# Remove the call to ray.wait if you want evaluation to run
# in parallel with training
ray.wait([
evaluate.remote(eval_agent_spec, evaluation_scenarios,
headless, seed)
])
env.close()
def test_graceful_shutdown():
"""SMARTS should not throw any exceptions when shutdown."""
agent_spec = AgentSpec(
interface=AgentInterface.from_type(AgentType.Laner),
agent_builder=lambda: Agent.from_function(lambda _: "keep_lane"),
)
env = build_env(agent_spec)
agent = agent_spec.build_agent()
obs = env.reset()
for _ in range(10):
obs = env.step({AGENT_ID: agent.act(obs)})
env.close()
def entrypoint(
goal_is_nearby_threshold=40,
lane_end_threshold=51,
lane_crash_distance_threshold=6,
lane_crash_ttc_threshold=2,
intersection_crash_distance_threshold=6,
intersection_crash_ttc_threshold=5,
target_speed=15,
lane_change_speed=12.5,
):
with pkg_resources.path(rl_agent, "checkpoint") as checkpoint_path:
return AgentSpec(
interface=agent_interface,
observation_adapter=get_observation_adapter(
goal_is_nearby_threshold=goal_is_nearby_threshold,
lane_end_threshold=lane_end_threshold,
lane_crash_distance_threshold=lane_crash_distance_threshold,
lane_crash_ttc_threshold=lane_crash_ttc_threshold,
intersection_crash_distance_threshold=
intersection_crash_distance_threshold,
intersection_crash_ttc_threshold=
intersection_crash_ttc_threshold,
),
action_adapter=get_action_adapter(
target_speed=target_speed,
lane_change_speed=lane_change_speed,
),
agent_builder=lambda: RLAgent(
load_path=str((checkpoint_path / "checkpoint").absolute()),
policy_name="default_policy",
observation_space=OBSERVATION_SPACE,
action_space=ACTION_SPACE,
),
)
def _make_agent_specs(num_agent):
agent_specs = {
"AGENT_" + str(agent_id): AgentSpec(
interface=AgentInterface(
rgb=RGB(),
action=ActionSpaceType.Lane,
),
agent_builder=lambda: Agent.from_function(lambda _: "keep_lane"),
observation_adapter=lambda obs: obs.top_down_rgb.data,
reward_adapter=lambda obs, reward: reward,
info_adapter=lambda obs, reward, info: info["score"],
)
for agent_id in range(num_agent)
}
obs_space = gym.spaces.Dict({
"AGENT_" + str(agent_id): gym.spaces.Box(
low=0,
high=255,
shape=(
agent_specs["AGENT_" + str(agent_id)].interface.rgb.width,
agent_specs["AGENT_" + str(agent_id)].interface.rgb.height,
3,
),
dtype=np.uint8,
)
for agent_id in range(num_agent)
})
return agent_specs, obs_space
def main(scenarios, headless, seed):
agent_spec = AgentSpec(
interface=AgentInterface.from_type(AgentType.Laner,
max_episode_steps=None),
agent_builder=None,
observation_adapter=None,
)
smarts = SMARTS(
agent_interfaces={},
traffic_sim=SumoTrafficSimulation(headless=True, auto_start=True),
envision=Envision(),
)
scenarios_iterator = Scenario.scenario_variations(
scenarios,
list([]),
)
smarts.reset(next(scenarios_iterator))
for _ in range(5000):
smarts.step({})
smarts.attach_sensors_to_vehicles(
agent_spec, smarts.vehicle_index.social_vehicle_ids())
obs, _, _, _ = smarts.observe_from(
smarts.vehicle_index.social_vehicle_ids())
def agent_spec():
return AgentSpec(
interface=AgentInterface.from_type(AgentType.TrajectoryInterpolator,
neighborhood_vehicles=True),
agent_builder=WithTimeTrajectoryAgent,
agent_params=None,
)
def replay_entrypoint(
save_directory,
id,
wrapped_agent_locator,
wrapped_agent_params=None,
read=False,
):
if wrapped_agent_params is None:
wrapped_agent_params = {}
from .replay_agent import ReplayAgent
internal_spec = make(wrapped_agent_locator, **wrapped_agent_params)
global social_index
global replay_save_dir
global replay_read
spec = AgentSpec(
interface=internal_spec.interface,
agent_params={
"save_directory": replay_save_dir,
"id": f"{id}_{social_index}",
"internal_spec": internal_spec,
"wrapped_agent_params": wrapped_agent_params,
"read": replay_read,
},
agent_builder=ReplayAgent,
)
social_index += 1
return spec
def agent_spec(agent_and_agent_type):
return AgentSpec(
interface=AgentInterface.from_type(
agent_and_agent_type[1], max_episode_steps=5000
),
agent_builder=agent_and_agent_type[0],
)
def __init__(self, save_directory, id, read: bool,
internal_spec: AgentSpec):
import smarts.core
if smarts.core.current_seed() is None:
smarts.core.seed(42)
self.save_directory = save_directory
self._base_agent = internal_spec.build_agent()
self._logger = logging.getLogger(self.__class__.__name__)
global agent_index
self.id = f"{id}_{agent_index}"
agent_index += 1
abs_path = os.path.abspath(save_directory)
self._read = read
file_mode = "wb" if not read else "rb"
path = Path(f"{abs_path}/{self.id}")
os.makedirs(abs_path, exist_ok=True)
try:
self._file = path.open(mode=file_mode)
except FileNotFoundError as e:
assert self._read
self._logger.error(
f"The file which you are trying to be read does not exist. "
f"Make sure the {save_directory} directory passed is correct and has the agent file which is being read"
)
raise e
def __init__(self, **kwargs):
print(kwargs)
self.episode_limit = kwargs['episode_limit']
self.n_agents = kwargs['agent_num']
self.observation_space = [
gym.spaces.Box(low=-1e10, high=1e10, shape=(10, ))
] * self.n_agents
self.action_space = [gym.spaces.Discrete(4)] * self.n_agents
self.agent_ids = ["Agent %i" % i for i in range(self.n_agents)]
self.n_actions = 4
self.scenarios = [kwargs['scenarios']]
self.headless = kwargs['headless']
num_episodes = 100
self.seed = kwargs['seed']
self.agent_specs = {
agent_id: AgentSpec(
interface=AgentInterface.from_type(AgentType.Laner,
max_episode_steps=5000),
observation_adapter=observation_adapter,
reward_adapter=reward_adapter,
action_adapter=action_adapter,
)
for agent_id in self.agent_ids
}
self.base_env = gym.make(
"smarts.env:hiway-v0",
scenarios=self.scenarios,
agent_specs=self.agent_specs,
headless=self.headless,
seed=self.seed,
)
self.current_observations = self.base_env.reset()
def entrypoint(
gains={
"theta": 3.0,
"position": 4.0,
"obstacle": 3.0,
"u_accel": 0.1,
"u_yaw_rate": 1.0,
"terminal": 0.01,
"impatience": 0.01,
"speed": 0.01,
},
debug=False,
max_episode_steps=600,
):
from .policy import Policy
return AgentSpec(
interface=AgentInterface(
action=ActionSpaceType.Trajectory,
waypoints=True,
neighborhood_vehicles=True,
max_episode_steps=max_episode_steps,
),
policy_params={"gains": gains, "debug": debug,},
policy_builder=Policy,
perform_self_test=False,
)
def entrypoint(
gains={
"theta": 3.0,
"position": 4.0,
"obstacle": 3.0,
"u_accel": 0.1,
"u_yaw_rate": 1.0,
"terminal": 0.01,
"impatience": 0.01,
"speed": 0.01,
"rate": 1,
},
debug=False,
aggressiveness=0,
max_episode_steps=None,
):
from .agent import OpEnAgent
return AgentSpec(
interface=AgentInterface(
action=ActionSpaceType.Trajectory,
waypoints=True,
neighborhood_vehicles=True,
max_episode_steps=max_episode_steps,
agent_behavior=AgentBehavior(aggressiveness=aggressiveness),
),
agent_params={
"gains": gains,
"debug": debug,
},
agent_builder=OpEnAgent,
)
def env_and_spec(action,
agent_type,
max_episode_steps,
scenarios,
seed=42,
agent_id="Agent-006"):
class Policy(AgentPolicy):
def act(self, obs):
return action
agent_spec = AgentSpec(
interface=AgentInterface.from_type(
agent_type, max_episode_steps=max_episode_steps),
policy_builder=Policy,
)
env = gym.make(
"smarts.env:hiway-v0",
scenarios=scenarios,
agent_specs={agent_id: agent_spec},
headless=True,
visdom=False,
timestep_sec=TIMESTEP_SEC,
sumo_headless=True,
seed=seed,
)
return (env, agent_spec)
def main(scenario):
scenario_path = Path(scenario).absolute()
agent_mission_count = Scenario.discover_agent_missions_count(scenario_path)
assert agent_mission_count > 0, "agent mission count should larger than 0"
agent_ids = [f"AGENT-{i}" for i in range(agent_mission_count)]
agent_specs = {
agent_id: AgentSpec(
interface=AgentInterface.from_type(AgentType.Laner, max_episode_steps=None),
agent_builder=RuleBasedAgent,
)
for agent_id in agent_ids
}
agents = {aid: agent_spec.build_agent() for aid, agent_spec in agent_specs.items()}
env = HiWayEnv(scenarios=[scenario_path], agent_specs=agent_specs)
while True:
observations = env.reset()
done = False
while not done:
agent_ids = list(observations.keys())
actions = {aid: agents[aid].act(observations[aid]) for aid in agent_ids}
observations, _, dones, _ = env.step(actions)
done = dones["__all__"]
def init_env():
if True:
agent_spec = AgentSpec(
interface=AgentInterface.from_type(
AgentType.Laner, max_episode_steps=all_args.episode_length
)
)
AGENT_ID = [str(i) for i in range(all_args.num_agents)]
env = gym.make(
"smarts.env:hiway-v0",
scenarios=all_args.scenarios,
agent_specs={i: agent_spec for i in AGENT_ID},
headless=all_args.headless,
visdom=False,
timestep_sec=0.1,
sumo_headless=True,
seed=all_args.seed + rank * 1000,
# zoo_workers=[("143.110.210.157", 7432)], # Distribute social agents across these workers
auth_key=all_args.auth_key,
# envision_record_data_replay_path="./data_replay",
)
env = SmartWrapper(env, all_args.num_agents)
else:
print("Can not support the " +
all_args.env_name + "environment.")
raise NotImplementedError
# env.seed(all_args.seed + rank * 1000)
return env
def main(_args):
scenario_path = Path(args.scenario).absolute()
mission_num = get_submission_num(scenario_path)
if mission_num == -1:
mission_num = 1
AGENT_IDS = [f"AGENT-{i}" for i in range(mission_num)]
agent_interface = AgentInterface.from_type(AgentType.Laner)
agent_specs = [
AgentSpec(interface=agent_interface,
policy_builder=lambda: KeeplanePolicy())
for _ in range(mission_num)
]
agents = dict(zip(AGENT_IDS, agent_specs))
env = gym.make(
"smarts.env:hiway-v0",
scenarios=[scenario_path],
agent_specs=agents,
headless=_args.headless,
visdom=False,
seed=42,
)
agents = {
_id: agent_spec.build_agent()
for _id, agent_spec in agents.items()
}
import webbrowser
webbrowser.open('http://localhost:8081/')
for ie in range(30):
step = 0
print(f"\n---- Starting episode: {ie}...")
observations = env.reset()
total_reward = 0.0
dones = {"__all__": False}
while not dones["__all__"]:
step += 1
agent_actions = {
_id: agents[_id].act(obs)
for _id, obs in observations.items()
}
observations, rewards, dones, _ = env.step(agent_actions)
total_reward += sum(rewards.values())
if (step + 1) % 10 == 0:
print(
f"* Episode: {ie} * step: {step} * acc-Reward: {total_reward}"
)
print("Accumulated reward:", total_reward)
env.close()
def run_experiment(log_path, experiment_name, training_iteration=100):
model_path = Path(__file__).parent / "model"
agent_spec = AgentSpec(
interface=AgentInterface.from_type(AgentType.Standard,
max_episode_steps=5000),
policy=RLlibTFSavedModelAgent(
model_path.absolute(),
OBSERVATION_SPACE,
),
observation_adapter=observation_adapter,
reward_adapter=reward_adapter,
action_adapter=action_adapter,
)
rllib_policies = {
"policy": (
None,
OBSERVATION_SPACE,
ACTION_SPACE,
{
"model": {
"custom_model": TrainingModel.NAME
}
},
)
}
scenario_path = Path(__file__).parent / "../../scenarios/loop"
scenario_path = str(scenario_path.absolute())
tune_confg = {
"env": RLlibHiWayEnv,
"env_config": {
"scenarios": [scenario_path],
"seed": 42,
"headless": True,
"agent_specs": {
"Agent-007": agent_spec
},
},
"multiagent": {
"policies": rllib_policies,
"policy_mapping_fn": lambda _: "policy",
},
"log_level": "WARN",
"num_workers": multiprocessing.cpu_count() - 1,
"horizon": HORIZON,
}
analysis = tune.run(
"PPO",
name=experiment_name,
stop={"training_iteration": training_iteration},
max_failures=10,
local_dir=log_path,
config=tune_confg,
)
return analysis
def rllib_agent():
def observation_adapter(env_observation):
ego = env_observation.ego_vehicle_state
waypoint_paths = env_observation.waypoint_paths
wps = [path[0] for path in waypoint_paths]
# distance of vehicle from center of lane
closest_wp = min(wps, key=lambda wp: wp.dist_to(ego.position))
signed_dist_from_center = closest_wp.signed_lateral_error(ego.position)
lane_hwidth = closest_wp.lane_width * 0.5
norm_dist_from_center = signed_dist_from_center / lane_hwidth
return {
"distance_from_center": np.array([norm_dist_from_center]),
"angle_error": np.array([closest_wp.relative_heading(ego.heading)]),
"speed": np.array([ego.speed]),
"steering": np.array([ego.steering]),
}
def reward_adapter(env_obs, env_reward):
return env_reward
def action_adapter(model_action):
throttle, brake, steering = model_action
return np.array([throttle, brake, steering])
def info_adapter(env_obs, env_reward, env_info):
env_info[INFO_EXTRA_KEY] = "blah"
return env_info
# This action space should match the input to the action_adapter(..) function below.
ACTION_SPACE = gym.spaces.Box(
low=np.array([0.0, 0.0, -1.0]), high=np.array([1.0, 1.0, 1.0]), dtype=np.float32
)
# This observation space should match the output of observation_adapter(..) below
OBSERVATION_SPACE = gym.spaces.Dict(
{
"distance_from_center": gym.spaces.Box(low=-1e10, high=1e10, shape=(1,)),
"angle_error": gym.spaces.Box(low=-np.pi, high=np.pi, shape=(1,)),
"speed": gym.spaces.Box(low=-1e10, high=1e10, shape=(1,)),
"steering": gym.spaces.Box(low=-1e10, high=1e10, shape=(1,)),
}
)
return {
"agent_spec": AgentSpec(
interface=AgentInterface.from_type(
AgentType.Standard, max_episode_steps=500
),
observation_adapter=observation_adapter,
reward_adapter=reward_adapter,
action_adapter=action_adapter,
info_adapter=info_adapter,
),
"observation_space": OBSERVATION_SPACE,
"action_space": ACTION_SPACE,
}
def main(scenarios, headless, seed):
scenarios_iterator = Scenario.scenario_variations(scenarios, [])
smarts = SMARTS(
agent_interfaces={},
traffic_sim=None,
envision=None if headless else Envision(),
)
for _ in scenarios:
scenario = next(scenarios_iterator)
agent_missions = scenario.discover_missions_of_traffic_histories()
for agent_id, mission in agent_missions.items():
agent_spec = AgentSpec(
interface=AgentInterface.from_type(AgentType.LanerWithSpeed,
max_episode_steps=None),
agent_builder=KeepLaneAgent,
agent_params=scenario.traffic_history_target_speed,
)
agent = agent_spec.build_agent()
# Take control of vehicle with corresponding agent_id
smarts.switch_ego_agent({agent_id: agent_spec.interface})
# tell the traffic history provider to start traffic
# at the point when this agent enters...
traffic_history_provider = smarts.get_provider_by_type(
TrafficHistoryProvider)
assert traffic_history_provider
traffic_history_provider.start_time = mission.start_time
# agent vehicle will enter right away...
modified_mission = replace(mission, start_time=0.0)
scenario.set_ego_missions({agent_id: modified_mission})
observations = smarts.reset(scenario)
dones = {agent_id: False}
while not dones.get(agent_id, True):
agent_obs = observations[agent_id]
agent_action = agent.act(agent_obs)
observations, rewards, dones, infos = smarts.step(
{agent_id: agent_action})
smarts.destroy()
def __new__(
self,
policy_class,
action_type,
checkpoint_dir=None,
task=None,
max_episode_steps=1200,
experiment_dir=None,
):
if experiment_dir:
print(f"LOADING SPEC from {experiment_dir}/spec.pkl")
with open(f"{experiment_dir}/spec.pkl", "rb") as input:
spec = dill.load(input)
new_spec = AgentSpec(
interface=spec.interface,
agent_params=dict(
policy_params=spec.agent_params["policy_params"],
checkpoint_dir=checkpoint_dir,
),
agent_builder=spec.policy_builder,
observation_adapter=spec.observation_adapter,
reward_adapter=spec.reward_adapter,
)
spec = new_spec
else:
adapter = BaselineAdapter()
policy_dir = "/".join(
inspect.getfile(policy_class).split("/")[:-1])
policy_params = load_yaml(f"{policy_dir}/params.yaml")
spec = AgentSpec(
interface=AgentInterface(
waypoints=Waypoints(lookahead=20),
neighborhood_vehicles=NeighborhoodVehicles(200),
action=action_type,
rgb=False,
max_episode_steps=max_episode_steps,
debug=True,
),
agent_params=dict(policy_params=policy_params,
checkpoint_dir=checkpoint_dir),
agent_builder=policy_class,
observation_adapter=adapter.observation_adapter,
reward_adapter=adapter.reward_adapter,
)
return spec
def on_trigger(ctx: Dict[str, Any]):
# Define agent specs to be assigned
agent_spec = AgentSpec(
interface=AgentInterface(waypoints=True, action=ActionSpaceType.Lane),
agent_builder=BasicAgent,
)
# Select a random sample from candidates
k = ctx.get("vehicles_to_replace_randomly", 0)
if k <= 0:
logger.warning(
"default (0) or negative value specified for replacement. Replacing all valid vehicle candidates."
)
sample = ctx["vehicle_candidates"]
else:
logger.info(
f"Choosing {k} vehicles randomly from {len(ctx['vehicle_candidates'])} valid vehicle candidates."
)
sample = random.sample(ctx["vehicle_candidates"], k)
assert len(sample) != 0
for veh_id in sample:
# Map selected vehicles to agent ids & specs
agent_id = f"agent-{veh_id}"
ctx["agents"][agent_id] = agent_spec.build_agent()
# Create missions based on current state and traffic history
positional, traverse = scenario.create_dynamic_traffic_history_mission(
veh_id, ctx["elapsed_sim_time"], ctx["positional_radius"]
)
# Take control of vehicles immediately
try:
# Try to assign a PositionalGoal at the last recorded timestep
smarts.add_agent_and_switch_control(
veh_id, agent_id, agent_spec.interface, positional
)
except PlanningError:
logger.warning(
f"Unable to create PositionalGoal for vehicle {veh_id}, falling back to TraverseGoal"
)
smarts.add_agent_and_switch_control(
veh_id, agent_id, agent_spec.interface, traverse
)