def main():
parser = argparse.ArgumentParser()
parser.add_argument('--env_name',
type=str,
default='maze2d-umaze-v0',
help='Maze type. small or default')
parser.add_argument('--num_episodes',
type=int,
default=100,
help='Num samples to collect')
args = parser.parse_args()
env = gym.make(args.env_name)
env.seed(0)
np.random.seed(0)
controller = waypoint_controller.WaypointController(env.str_maze_spec)
ravg = []
for _ in range(args.num_episodes):
s = env.reset()
returns = 0
for t in range(env._max_episode_steps):
position = s[0:2]
velocity = s[2:4]
act, done = controller.get_action(position, velocity,
env.get_target())
s, rew, _, _ = env.step(act)
returns += rew
ravg.append(returns)
print(args.env_name, 'returns', np.mean(ravg))
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--env_name',
type=str,
default='maze2d-umaze-v0',
help='Maze type. small or default')
parser.add_argument('--num_episodes',
type=int,
default=100,
help='Num samples to collect')
parser.add_argument('--render', action='store_true')
args = parser.parse_args()
env = gym.make(args.env_name)
if args.render:
env.render('human')
env.seed(0)
np.random.seed(0)
d_gain = -2.0
p_gain = 10.0
controller = waypoint_controller.WaypointController(env.env.str_maze_spec,
p_gain=p_gain,
d_gain=d_gain)
print('max steps:', env._max_episode_steps)
ravg = []
for _ in range(args.num_episodes):
controller = waypoint_controller.WaypointController(
env.env.str_maze_spec, p_gain=p_gain, d_gain=d_gain)
s = env.reset()
returns = 0
for t in range(env._max_episode_steps):
position = s[0:2]
velocity = s[2:4]
act, done = controller.get_action(position, velocity,
np.array(env.env.get_target()))
#print(position-1, controller.current_waypoint(), np.array(env.env.get_target()) - 1)
#print('\t', act)
s, rew, _, _ = env.step(act)
if args.render:
time.sleep(0.01)
env.render('human')
returns += rew
print(returns)
ravg.append(returns)
print(args.env_name, 'returns', np.mean(ravg))
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--render', action='store_true', help='Render trajectories')
parser.add_argument('--noisy', action='store_true', help='Noisy actions')
parser.add_argument('--env_name', type=str, default='maze2d-umaze-v1', help='Maze type')
parser.add_argument('--num_samples', type=int, default=int(1e6), help='Num samples to collect')
args = parser.parse_args()
env = gym.make(args.env_name)
maze = env.str_maze_spec
max_episode_steps = env._max_episode_steps
# default: p=10, d=-1
controller = waypoint_controller.WaypointController(maze, p_gain=10.0, d_gain=-2.0)
env = bullet_maze.Maze2DBulletEnv(maze)
if args.render:
env.render('human')
env.set_target()
s = env.reset()
act = env.action_space.sample()
timeout = False
data = reset_data()
last_position = s[0:2]
ts = 0
for _ in range(args.num_samples):
position = s[0:2]
velocity = s[2:4]
# subtract 1.0 due to offset between tabular maze representation and bullet state
act, done = controller.get_action(position , velocity, env._target)
if args.noisy:
act = act + np.random.randn(*act.shape)*0.5
act = np.clip(act, -1.0, 1.0)
if ts >= max_episode_steps:
timeout = True
append_data(data, s, act, env._target, done, timeout, env.robot)
ns, _, _, _ = env.step(act)
if len(data['observations']) % 10000 == 0:
print(len(data['observations']))
ts += 1
if done:
env.set_target()
done = False
ts = 0
else:
last_position = s[0:2]
s = ns
if args.render:
env.render('human')
if args.noisy:
fname = '%s-noisy-bullet.hdf5' % args.env_name
else:
fname = '%s-bullet.hdf5' % args.env_name
dataset = h5py.File(fname, 'w')
npify(data)
for k in data:
dataset.create_dataset(k, data=data[k], compression='gzip')
def sample_env_and_controller(args, layout):
layout_str = semantic_layout2str(layout)
env = maze_model.MazeEnv(layout_str, agent_centric_view=args.agent_centric)
controller = waypoint_controller.WaypointController(layout_str)
return env, controller
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--render',
action='store_true',
help='Render trajectories')
parser.add_argument('--noisy', action='store_true', help='Noisy actions')
parser.add_argument('--env_name',
type=str,
default='maze2d-umaze-v1',
help='Maze type')
parser.add_argument('--num_samples',
type=int,
default=int(1e6),
help='Num samples to collect')
args = parser.parse_args()
env = gym.make(args.env_name)
maze = env.str_maze_spec
max_episode_steps = env._max_episode_steps
controller = waypoint_controller.WaypointController(maze)
env = maze_model.MazeEnv(maze)
env.set_target()
s = env.reset()
act = env.action_space.sample()
done = False
data = reset_data()
ts = 0
for _ in range(args.num_samples):
position = s[0:2]
velocity = s[2:4]
act, done = controller.get_action(position, velocity, env._target)
if args.noisy:
act = act + np.random.randn(*act.shape) * 0.5
act = np.clip(act, -1.0, 1.0)
if ts >= max_episode_steps:
done = True
append_data(data, s, act, env._target, done, env.sim.data)
ns, _, _, _ = env.step(act)
if len(data['observations']) % 10000 == 0:
print(len(data['observations']))
ts += 1
if done:
env.set_target()
done = False
ts = 0
else:
s = ns
if args.render:
env.render()
if args.noisy:
fname = '%s-noisy.hdf5' % args.env_name
else:
fname = '%s.hdf5' % args.env_name
dataset = h5py.File(fname, 'w')
npify(data)
for k in data:
dataset.create_dataset(k, data=data[k], compression='gzip')
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--render',
action='store_true',
help='Render trajectories')
parser.add_argument('--noisy', action='store_true', help='Noisy actions')
parser.add_argument('--agent_centric',
action='store_true',
help='Whether agent-centric images are rendered.')
parser.add_argument('--save_images',
action='store_true',
help='Whether rendered images are saved.')
parser.add_argument('--maze',
type=str,
default='hardexpv2',
help='Maze type. small or default')
parser.add_argument('--data_dir',
type=str,
default='.',
help='Base directory for dataset')
parser.add_argument('--num_samples',
type=int,
default=int(2e5),
help='Num samples to collect')
parser.add_argument('--min_traj_len',
type=int,
default=int(20),
help='Min number of samples per trajectory')
parser.add_argument('--batch_idx',
type=int,
default=int(-1),
help='(Optional) Index of generated data batch')
args = parser.parse_args()
if args.agent_centric and not args.save_images:
raise ValueError("Need to save images for agent-centric dataset")
if args.maze == 'umaze':
maze = maze_layouts.U_MAZE
max_episode_steps = 150
elif args.maze == 'open':
maze = maze_layouts.OPEN
max_episode_steps = 150
elif args.maze == 'medium':
maze = maze_layouts.MEDIUM_MAZE
max_episode_steps = 250
elif args.maze == 'hardexp':
maze = maze_layouts.HARD_EXP_MAZE
max_episode_steps = 300
elif args.maze == 'hardexpv2':
maze = maze_layouts.HARD_EXP_MAZE_V2
max_episode_steps = 1600 if not args.agent_centric else 400
else:
maze = maze_layouts.LARGE_MAZE
max_episode_steps = 600
controller = waypoint_controller.WaypointController(maze)
env = maze_model.MazeEnv(maze, agent_centric_view=args.agent_centric)
s = reset_env(env, agent_centric=args.agent_centric)
data = reset_data()
ts, cnt = 0, 0
for tt in tqdm.tqdm(range(args.num_samples)):
position = s[0:2]
velocity = s[2:4]
act, done = controller.get_action(position, velocity, env._target)
if args.noisy:
act = act + np.random.randn(*act.shape) * 0.5
act = np.clip(act, -1.0, 1.0)
if ts >= max_episode_steps:
done = True
append_data(data, s, act, env.render(mode='rgb_array'), env._target,
done, env.sim.data)
ns, _, _, _ = env.step(act)
ts += 1
if done:
if len(data['actions']) > args.min_traj_len:
save_data(args, data, cnt)
cnt += 1
data = reset_data()
s = reset_env(env, agent_centric=args.agent_centric)
ts = 0
else:
s = ns
if args.render:
env.render(mode='human')
def sample_env_and_controller(args):
layout_str = maze_layouts.rand_layout(seed=0, size=args.fixed_maze_size)
env = maze_model.MazeEnv(layout_str, agent_centric_view=args.agent_centric)
controller = waypoint_controller.WaypointController(layout_str)
return env, controller
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--render', action='store_true', help='Render trajectories')
parser.add_argument('--noisy', action='store_true', help='Noisy actions')
parser.add_argument('--env_name', type=str, default='maze2d-umaze-v1', help='Maze type')
parser.add_argument('--num_samples', type=int, default=int(1e6), help='Num samples to collect')
parser.add_argument('--data_dir', type=str, default='.', help='Base directory for dataset')
parser.add_argument('--batch_idx', type=int, default=int(-1), help='(Optional) Index of generated data batch')
args = parser.parse_args()
env = gym.make(args.env_name)
maze = env.str_maze_spec
max_episode_steps = env._max_episode_steps
controller = waypoint_controller.WaypointController(maze)
env = maze_model.MazeEnv(maze)
env.set_target()
s = env.reset()
act = env.action_space.sample()
done = False
data = reset_data()
ts = 0
for _ in range(args.num_samples):
position = s[0:2]
velocity = s[2:4]
act, done = controller.get_action(position, velocity, env._target)
if args.noisy:
act = act + np.random.randn(*act.shape)*0.5
act = np.clip(act, -1.0, 1.0)
if ts >= max_episode_steps:
done = True
append_data(data, s, act, env._target, done, env.sim.data)
ns, _, _, _ = env.step(act)
if len(data['observations']) % 1000 == 0:
print(len(data['observations']))
ts += 1
if done:
env.set_target()
done = False
ts = 0
else:
s = ns
if args.render:
env.render()
if args.batch_idx >= 0:
dir_name = 'maze2d-%s-noisy' % args.maze if args.noisy else 'maze2d-%s-sparse' % args.maze
os.makedirs(os.path.join(args.data_dir, dir_name), exist_ok=True)
fname = os.path.join(args.data_dir, dir_name, "rollouts_batch_{}.h5".format(args.batch_idx))
else:
os.makedirs(args.data_dir, exist_ok=True)
fname = 'maze2d-%s-noisy.hdf5' % args.maze if args.noisy else 'maze2d-%s-sparse.hdf5' % args.maze
fname = os.path.join(args.data_dir, fname)
dataset = h5py.File(fname, 'w')
npify(data)
for k in data:
dataset.create_dataset(k, data=data[k], compression='gzip')