def make_env(config, writer, prefix, datadir, store):
suite, task = config.task.split('_', 1)
if suite == 'kitchen':
env = wrappers.Kitchen2D(task)
env = wrappers.ActionRepeat(env, config.action_repeat)
env = wrappers.NormalizeActions(env)
elif suite == 'dmc':
env = wrappers.DeepMindControl(task)
env = wrappers.ActionRepeat(env, config.action_repeat)
env = wrappers.NormalizeActions(env)
elif suite == 'atari':
env = wrappers.Atari(
task, config.action_repeat, (64, 64), grayscale=False,
life_done=True, sticky_actions=True)
env = wrappers.OneHotAction(env)
else:
raise NotImplementedError(suite)
env = wrappers.TimeLimit(env, config.time_limit / config.action_repeat)
callbacks = []
if store:
callbacks.append(lambda ep: tools.save_episodes(datadir, [ep]))
if prefix == 'test':
callbacks.append(
lambda ep: summarize_episode(ep, config, datadir, writer, prefix))
env = wrappers.Collect(env, callbacks, config.precision)
env = wrappers.RewardObs(env)
return env
def make_env(config, writer, prefix, datadir, store):
suite, task = config.task.split('_', 1)
if suite == 'dmc':
env = wrappers.DeepMindControl(task)
env = wrappers.ActionRepeat(env, config.action_repeat)
env = wrappers.NormalizeActions(env)
elif suite == 'gym':
env = wrappers.Gym(task, config, size=(128, 128))
env = wrappers.ActionRepeat(env, config.action_repeat)
env = wrappers.NormalizeActions(env)
elif task == 'door':
env = wrappers.DoorOpen(config, size=(128, 128))
env = wrappers.ActionRepeat(env, config.action_repeat)
env = wrappers.NormalizeActions(env)
elif task == 'drawer':
env = wrappers.DrawerOpen(config, size=(128, 128))
env = wrappers.ActionRepeat(env, config.action_repeat)
env = wrappers.NormalizeActions(env)
else:
raise NotImplementedError(suite)
env = wrappers.TimeLimit(env, config.time_limit / config.action_repeat)
callbacks = []
if store:
callbacks.append(lambda ep: tools.save_episodes(datadir, [ep]))
if prefix == 'test':
callbacks.append(
lambda ep: summarize_episode(ep, config, datadir, writer, prefix))
env = wrappers.Collect(env, callbacks, config.precision)
env = wrappers.RewardObs(env)
return env
def make_env(config,
writer,
prefix,
datadir,
store,
index=None,
real_world=False):
suite, task = config.task.split('_', 1)
if suite == 'dmc':
if config.dr is None or real_world: #first index is always real world
env = wrappers.DeepMindControl(task,
use_state=config.use_state,
real_world=real_world)
else:
env = wrappers.DeepMindControl(task,
dr=config.dr,
use_state=config.use_state,
real_world=real_world)
env = wrappers.ActionRepeat(env, config.action_repeat)
env = wrappers.NormalizeActions(env)
elif suite == 'atari':
env = wrappers.Atari(task,
config.action_repeat, (64, 64),
grayscale=False,
life_done=True,
sticky_actions=True)
env = wrappers.OneHotAction(env)
elif suite == 'gym':
if index == 0 or index is None: #first index is always real world
env = wrappers.GymControl(task)
else:
env = wrappers.GymControl(task, dr=config.dr)
env = wrappers.ActionRepeat(env, config.action_repeat)
env = wrappers.NormalizeActions(env)
else:
raise NotImplementedError(suite)
env = wrappers.TimeLimit(env, config.time_limit / config.action_repeat)
callbacks = []
if store:
callbacks.append(lambda ep: tools.save_episodes(datadir, [ep]))
callbacks.append(
lambda ep: summarize_episode(ep, config, datadir, writer, prefix))
env = wrappers.Collect(env, callbacks, config.precision)
env = wrappers.RewardObs(env)
return env
def make_base_env(config, gui=False):
env = wrappers.RaceCarBaseEnv(track=config.track,
task=config.task,
rendering=gui)
env = wrappers.RaceCarWrapper(env, agent_id='A')
env = wrappers.ActionRepeat(env, config.action_repeat)
env = wrappers.ReduceActionSpace(env, low=[0.005, -1.0], high=[1.0, 1.0])
env = wrappers.OccupancyMapObs(env)
return env
def make_single_track_env(track, action_repeat, rendering=True):
scenario = MultiAgentScenario.from_spec(f'scenarios/eval/{track}.yml',
rendering=rendering)
env = MultiAgentRaceEnv(scenario=scenario)
env = wrappers.RaceCarWrapper(env)
env = wrappers.FixedResetMode(env, mode='grid')
env = wrappers.ActionRepeat(env, action_repeat)
env = wrappers.ReduceActionSpace(env, low=[0.005, -1.0], high=[1.0, 1.0])
env = wrappers.OccupancyMapObs(env)
return env
def make_multi_track_env(tracks,
action_repeat,
rendering=True,
is_dreamer=True):
# note: problem of multi-track racing env with wrapper `OccupancyMapObs` because it initializes the map once
# ideas to solve this issue? when changing env force the update of occupancy map in wrapper?
scenarios = [
MultiAgentScenario.from_spec(f'scenarios/eval/{track}.yml',
rendering=rendering) for track in tracks
]
env = ChangingTrackMultiAgentRaceEnv(scenarios=scenarios, order='manual')
env = wrappers.RaceCarWrapper(env)
env = wrappers.FixedResetMode(env, mode='grid')
env = wrappers.ActionRepeat(env, action_repeat)
if is_dreamer:
env = wrappers.ReduceActionSpace(env,
low=[0.005, -1.0],
high=[1.0, 1.0])
return env
def make_env(config, writer, prefix, datadir, store):
suite, task = config.task.split('_', 1)
if suite == 'dmc':
env = wrappers.DeepMindControl(task)
env = wrappers.ActionRepeat(env, config.action_repeat)
env = wrappers.NormalizeActions(env)
elif suite == 'atari':
env = wrappers.Atari(task,
config.action_repeat, (64, 64),
grayscale=False,
life_done=True,
sticky_actions=True)
env = wrappers.OneHotAction(env)
elif suite == 'football':
env = football_env.create_environment(
representation='pixels',
env_name='academy_empty_goal_close',
stacked=False,
logdir='./football/empty_goal_close2',
write_goal_dumps=True,
write_full_episode_dumps=True,
render=True,
write_video=True)
env = wrappers.Football(env)
env = wrappers.OneHotAction(env)
else:
raise NotImplementedError(suite)
env = wrappers.TimeLimit(env, config.time_limit / config.action_repeat)
callbacks = []
if store:
callbacks.append(lambda ep: tools.save_episodes(datadir, [ep]))
callbacks.append(
lambda ep: summarize_episode(ep, config, datadir, writer, prefix))
env = wrappers.Collect(env, callbacks, config.precision)
env = wrappers.RewardObs(env)
return env
import gym
import wrappers
import numpy as np
task = 'SingleAgentTreitlstrasse_v2-v0'
time_limit = 60 * 100
action_repeat = 8
env = gym.make(task)
env = wrappers.TimeLimit(env, time_limit)
env = wrappers.ActionRepeat(env, action_repeat)
def test_on_track(model, outdir):
video, returns = simulate_episode(model)
videodir = outdir / 'videos'
videodir.mkdir(parents=True, exist_ok=True)
import imageio
writer = imageio.get_writer(videodir / f'test_return{returns}.mp4')
for image in video:
writer.append_data(image)
writer.close()
def simulate_episode(model, prediction_window=5, terminate_on_collision=True):
# to do: make it uniform to f1_tenth directory
done = False
obs = env.reset(mode='grid')
state = None
video = []
returns = 0.0
while not done:
