def __init__(
self,
device_id=-1,
headless=False,
render=True,
horizon=100,
camera_width=240, #used to be 720
camera_height=240,
num_variations=1,
action_repeat=8,
camera_name='default_camera',
deterministic=True,
use_cached_states=True,
save_cached_states=True,
**kwargs):
self.camera_params, self.camera_width, self.camera_height, self.camera_name = {}, camera_width, camera_height, camera_name
pyflex.init(headless, render, camera_width, camera_height)
self.record_video, self.video_path, self.video_name = False, None, None
self.metadata = {'render.modes': ['human', 'rgb_array']}
if device_id == -1 and 'gpu_id' in os.environ:
device_id = int(os.environ['gpu_id'])
self.device_id = device_id
self.horizon = horizon
self.time_step = 0
self.action_repeat = action_repeat
self.recording = False
self.prev_reward = None
self.deterministic = deterministic
self.use_cached_states = use_cached_states
self.save_cached_states = save_cached_states
self.current_config = self.get_default_config()
self.current_config_id = None
self.cached_configs, self.cached_init_states = None, None
self.num_variations = num_variations
self.dim_position = 4
self.dim_velocity = 3
self.dim_shape_state = 14
self.particle_num = 0
self.eval_flag = False
# version 1 does not support robot, while version 2 does.
pyflex_root = os.environ['PYFLEXROOT']
if 'Robotics' in pyflex_root:
self.version = 2
else:
self.version = 1
def gen_PyFleX(info):
env, env_idx = info['env'], info['env_idx']
thread_idx, data_dir, data_names = info['thread_idx'], info['data_dir'], info['data_names']
n_rollout, time_step = info['n_rollout'], info['time_step']
shape_state_dim, dt = info['shape_state_dim'], info['dt']
gen_vision = info['gen_vision']
vision_dir, vis_width, vis_height = info['vision_dir'], info['vis_width'], info['vis_height']
np.random.seed(round(time.time() * 1000 + thread_idx) % 2 ** 32)
# positions
stats = [init_stat(3)]
import pyflex
pyflex.init()
for i in range(n_rollout):
if i % 10 == 0:
print("%d / %d" % (i, n_rollout))
rollout_idx = thread_idx * n_rollout + i
rollout_dir = os.path.join(data_dir, str(rollout_idx))
os.system('mkdir -p ' + rollout_dir)
if env == 'RigidFall':
g_low, g_high = info['physics_param_range']
gravity = rand_float(g_low, g_high)
print("Generated RigidFall rollout {} with gravity {} from range {} ~ {}".format(
i, gravity, g_low, g_high))
n_instance = 3
draw_mesh = 1
scene_params = np.zeros(n_instance * 3 + 3)
scene_params[0] = n_instance
scene_params[1] = gravity
scene_params[-1] = draw_mesh
low_bound = 0.09
for j in range(n_instance):
x = rand_float(0., 0.1)
y = rand_float(low_bound, low_bound + 0.01)
z = rand_float(0., 0.1)
scene_params[j * 3 + 2] = x
scene_params[j * 3 + 3] = y
scene_params[j * 3 + 4] = z
low_bound += 0.21
pyflex.set_scene(env_idx, scene_params, thread_idx)
pyflex.set_camPos(np.array([0.2, 0.875, 2.0]))
n_particles = pyflex.get_n_particles()
n_shapes = 1 # the floor
positions = np.zeros((time_step, n_particles + n_shapes, 3), dtype=np.float32)
shape_quats = np.zeros((time_step, n_shapes, 4), dtype=np.float32)
for j in range(time_step):
positions[j, :n_particles] = pyflex.get_positions().reshape(-1, 4)[:, :3]
ref_positions = positions[0]
for k in range(n_instance):
XX = ref_positions[64*k:64*(k+1)]
YY = positions[j, 64*k:64*(k+1)]
X = XX.copy().T
Y = YY.copy().T
mean_X = np.mean(X, 1, keepdims=True)
mean_Y = np.mean(Y, 1, keepdims=True)
X = X - mean_X
Y = Y - mean_Y
C = np.dot(X, Y.T)
U, S, Vt = np.linalg.svd(C)
D = np.eye(3)
D[2, 2] = np.linalg.det(np.dot(Vt.T, U.T))
R = np.dot(Vt.T, np.dot(D, U.T))
t = mean_Y - np.dot(R, mean_X)
YY_fitted = (np.dot(R, XX.T) + t).T
# print("MSE fit", np.mean(np.square(YY_fitted - YY)))
positions[j, 64*k:64*(k+1)] = YY_fitted
if gen_vision:
pyflex.step(capture=True, path=os.path.join(rollout_dir, str(j) + '.tga'))
else:
pyflex.step()
data = [positions[j], shape_quats[j], scene_params]
store_data(data_names, data, os.path.join(rollout_dir, str(j) + '.h5'))
if gen_vision:
images = np.zeros((time_step, vis_height, vis_width, 3), dtype=np.uint8)
for j in range(time_step):
img_path = os.path.join(rollout_dir, str(j) + '.tga')
img = scipy.misc.imread(img_path)[:, :, :3][:, :, ::-1]
img = cv2.resize(img, (vis_width, vis_height), interpolation=cv2.INTER_AREA)
images[j] = img
os.system('rm ' + img_path)
store_data(['positions', 'images', 'scene_params'], [positions, images, scene_params],
os.path.join(vision_dir, str(rollout_idx) + '.h5'))
elif env == 'MassRope':
s_low, s_high = info['physics_param_range']
stiffness = rand_float(s_low, s_high)
print("Generated MassRope rollout {} with gravity {} from range {} ~ {}".format(
i, stiffness, s_low, s_high))
x = 0.
y = 1.0
z = 0.
length = 0.7
draw_mesh = 1.
scene_params = np.array([x, y, z, length, stiffness, draw_mesh])
pyflex.set_scene(env_idx, scene_params, 0)
pyflex.set_camPos(np.array([0.13, 2.0, 3.2]))
action = np.zeros(3)
# the last particle is the pin, regarded as shape
n_particles = pyflex.get_n_particles() - 1
n_shapes = 1 # the mass at the top of the rope
positions = np.zeros((time_step + 1, n_particles + n_shapes, 3), dtype=np.float32)
shape_quats = np.zeros((time_step + 1, n_shapes, 4), dtype=np.float32)
action = np.zeros(3)
for j in range(time_step + 1):
positions[j] = pyflex.get_positions().reshape(-1, 4)[:, :3]
if j >= 1:
# append the action (position of the pin) to the previous time step
positions[j - 1, -1, :] = positions[j, -1, :]
ref_positions = positions[0]
# apply rigid projection to the rigid object
# cube: [0, 81)
# rope: [81, 95)
# pin: [95, 96)
XX = ref_positions[:81]
YY = positions[j, :81]
X = XX.copy().T
Y = YY.copy().T
mean_X = np.mean(X, 1, keepdims=True)
mean_Y = np.mean(Y, 1, keepdims=True)
X = X - mean_X
Y = Y - mean_Y
C = np.dot(X, Y.T)
U, S, Vt = np.linalg.svd(C)
D = np.eye(3)
D[2, 2] = np.linalg.det(np.dot(Vt.T, U.T))
R = np.dot(Vt.T, np.dot(D, U.T))
t = mean_Y - np.dot(R, mean_X)
YY_fitted = (np.dot(R, XX.T) + t).T
positions[j, :81] = YY_fitted
scale = 0.1
action[0] += rand_float(-scale, scale) - positions[j, -1, 0] * 0.1
action[2] += rand_float(-scale, scale) - positions[j, -1, 2] * 0.1
if gen_vision:
pyflex.step(action * dt, capture=True, path=os.path.join(rollout_dir, str(j) + '.tga'))
else:
pyflex.step(action * dt)
if j >= 1:
data = [positions[j - 1], shape_quats[j - 1], scene_params]
store_data(data_names, data, os.path.join(rollout_dir, str(j - 1) + '.h5'))
if gen_vision:
images = np.zeros((time_step, vis_height, vis_width, 3), dtype=np.uint8)
for j in range(time_step):
img_path = os.path.join(rollout_dir, str(j) + '.tga')
img = scipy.misc.imread(img_path)[:, :, :3][:, :, ::-1]
img = cv2.resize(img, (vis_width, vis_height), interpolation=cv2.INTER_AREA)
images[j] = img
os.system('rm ' + img_path)
store_data(['positions', 'images', 'scene_params'], [positions, images, scene_params],
os.path.join(vision_dir, str(rollout_idx) + '.h5'))
else:
raise AssertionError("Unsupported env")
# change dtype for more accurate stat calculation
# only normalize positions
datas = [positions[:time_step].astype(np.float64)]
for j in range(len(stats)):
stat = init_stat(stats[j].shape[0])
stat[:, 0] = np.mean(datas[j], axis=(0, 1))[:]
stat[:, 1] = np.std(datas[j], axis=(0, 1))[:]
stat[:, 2] = datas[j].shape[0] * datas[j].shape[1]
stats[j] = combine_stat(stats[j], stat)
pyflex.clean()
return stats
[x_curr, (height + border) / 2., -(dis_z + border) / 2.])
states[3, 3:6] = np.array(
[x_last, (height + border) / 2., -(dis_z + border) / 2.])
states[4, :3] = np.array(
[x_curr, (height + border) / 2., (dis_z + border) / 2.])
states[4, 3:6] = np.array(
[x_last, (height + border) / 2., (dis_z + border) / 2.])
states[:, 6:10] = quat
states[:, 10:] = quat
return states
pyflex.init()
use_gpu = torch.cuda.is_available()
def rand_float(lo, hi):
return np.random.rand() * (hi - lo) + lo
def rand_int(lo, hi):
return np.random.randint(lo, hi)
### set scene
# dim_x: [8, 12]
# dim_y: [8, 12]
def gen_PyFleX(info):
env, root_num = info['env'], info['root_num']
thread_idx, data_dir, data_names = info['thread_idx'], info['data_dir'], info['data_names']
n_rollout, n_instance = info['n_rollout'], info['n_instance']
time_step, time_step_clip = info['time_step'], info['time_step_clip']
shape_state_dim, dt = info['shape_state_dim'], info['dt']
env_idx = info['env_idx']
np.random.seed(round(time.time() * 1000 + thread_idx) % 2**32) ### NOTE: we might want to fix the seed for reproduction
# positions, velocities
if env_idx == 5: # RiceGrip
stats = [init_stat(6), init_stat(6)]
else:
stats = [init_stat(3), init_stat(3)]
import pyflex
pyflex.init()
for i in range(n_rollout):
if i % 10 == 0:
print("%d / %d" % (i, n_rollout))
rollout_idx = thread_idx * n_rollout + i
rollout_dir = os.path.join(data_dir, str(rollout_idx))
os.system('mkdir -p ' + rollout_dir)
if env == 'FluidFall':
scene_params = np.zeros(1)
pyflex.set_scene(env_idx, scene_params, thread_idx)
n_particles = pyflex.get_n_particles()
positions = np.zeros((time_step, n_particles, 3), dtype=np.float32)
velocities = np.zeros((time_step, n_particles, 3), dtype=np.float32)
for j in range(time_step_clip):
p_clip = pyflex.get_positions().reshape(-1, 4)[:, :3]
pyflex.step()
for j in range(time_step):
positions[j] = pyflex.get_positions().reshape(-1, 4)[:, :3]
if j == 0:
velocities[j] = (positions[j] - p_clip) / dt
else:
velocities[j] = (positions[j] - positions[j - 1]) / dt
pyflex.step()
data = [positions[j], velocities[j]]
store_data(data_names, data, os.path.join(rollout_dir, str(j) + '.h5'))
elif env == 'BoxBath':
# BoxBath
scene_params = np.zeros(1)
pyflex.set_scene(env_idx, scene_params, thread_idx)
n_particles = pyflex.get_n_particles()
positions = np.zeros((time_step, n_particles, 3), dtype=np.float32)
velocities = np.zeros((time_step, n_particles, 3), dtype=np.float32)
for j in range(time_step_clip):
pyflex.step()
p = pyflex.get_positions().reshape(-1, 4)[:64, :3]
clusters = []
st_time = time.time()
kmeans = MiniBatchKMeans(n_clusters=root_num[0][0], random_state=0).fit(p)
# print('Time on kmeans', time.time() - st_time)
clusters.append([[kmeans.labels_]])
# centers = kmeans.cluster_centers_
ref_rigid = p
for j in range(time_step):
positions[j] = pyflex.get_positions().reshape(-1, 4)[:, :3]
# apply rigid projection to ground truth
XX = ref_rigid
YY = positions[j, :64]
# print("MSE init", np.mean(np.square(XX - YY)))
X = XX.copy().T
Y = YY.copy().T
mean_X = np.mean(X, 1, keepdims=True)
mean_Y = np.mean(Y, 1, keepdims=True)
X = X - mean_X
Y = Y - mean_Y
C = np.dot(X, Y.T)
U, S, Vt = np.linalg.svd(C)
D = np.eye(3)
D[2, 2] = np.linalg.det(np.dot(Vt.T, U.T))
R = np.dot(Vt.T, np.dot(D, U.T))
t = mean_Y - np.dot(R, mean_X)
YY_fitted = (np.dot(R, XX.T) + t).T
# print("MSE fit", np.mean(np.square(YY_fitted - YY)))
positions[j, :64] = YY_fitted
if j > 0:
velocities[j] = (positions[j] - positions[j - 1]) / dt
pyflex.step()
data = [positions[j], velocities[j], clusters]
store_data(data_names, data, os.path.join(rollout_dir, str(j) + '.h5'))
elif env == 'FluidShake':
# if env is FluidShake
height = 1.0
border = 0.025
dim_x = rand_int(10, 12)
dim_y = rand_int(15, 20)
dim_z = 3
x_center = rand_float(-0.2, 0.2)
x = x_center - (dim_x-1)/2.*0.055
y = 0.055/2. + border + 0.01
z = 0. - (dim_z-1)/2.*0.055
box_dis_x = dim_x * 0.055 + rand_float(0., 0.3)
box_dis_z = 0.2
scene_params = np.array([x, y, z, dim_x, dim_y, dim_z, box_dis_x, box_dis_z])
pyflex.set_scene(env_idx, scene_params, 0)
boxes = calc_box_init_FluidShake(box_dis_x, box_dis_z, height, border)
for i in range(len(boxes)):
halfEdge = boxes[i][0]
center = boxes[i][1]
quat = boxes[i][2]
pyflex.add_box(halfEdge, center, quat)
n_particles = pyflex.get_n_particles()
n_shapes = pyflex.get_n_shapes()
# print("n_particles", n_particles)
# print("n_shapes", n_shapes)
positions = np.zeros((time_step, n_particles + n_shapes, 3), dtype=np.float32)
velocities = np.zeros((time_step, n_particles + n_shapes, 3), dtype=np.float32)
shape_quats = np.zeros((time_step, n_shapes, 4), dtype=np.float32)
x_box = x_center
v_box = 0.
for j in range(time_step_clip):
x_box_last = x_box
x_box += v_box * dt
shape_states_ = calc_shape_states_FluidShake(
x_box, x_box_last, scene_params[-2:], height, border)
pyflex.set_shape_states(shape_states_)
pyflex.step()
for j in range(time_step):
x_box_last = x_box
x_box += v_box * dt
v_box += rand_float(-0.15, 0.15) - x_box * 0.1
shape_states_ = calc_shape_states_FluidShake(
x_box, x_box_last, scene_params[-2:], height, border)
pyflex.set_shape_states(shape_states_)
positions[j, :n_particles] = pyflex.get_positions().reshape(-1, 4)[:, :3]
shape_states = pyflex.get_shape_states().reshape(-1, shape_state_dim)
for k in range(n_shapes):
positions[j, n_particles + k] = shape_states[k, :3]
shape_quats[j, k] = shape_states[k, 6:10]
if j > 0:
velocities[j] = (positions[j] - positions[j - 1]) / dt
pyflex.step()
# NOTE: 1) particle + glass wall positions, 2) particle + glass wall velocitys, 3) glass wall rotations, 4) scenen parameters
data = [positions[j], velocities[j], shape_quats[j], scene_params]
store_data(data_names, data, os.path.join(rollout_dir, str(j) + '.h5'))
elif env == 'RiceGrip':
# if env is RiceGrip
# repeat the grip for R times
R = 3
gripper_config = sample_control_RiceGrip()
if i % R == 0:
### set scene
# x, y, z: [8.0, 10.0]
# clusterStiffness: [0.3, 0.7]
# clusterPlasticThreshold: [0.00001, 0.0005]
# clusterPlasticCreep: [0.1, 0.3]
x = rand_float(8.0, 10.0)
y = rand_float(8.0, 10.0)
z = rand_float(8.0, 10.0)
clusterStiffness = rand_float(0.3, 0.7)
clusterPlasticThreshold = rand_float(0.00001, 0.0005)
clusterPlasticCreep = rand_float(0.1, 0.3)
scene_params = np.array([x, y, z, clusterStiffness, clusterPlasticThreshold, clusterPlasticCreep])
pyflex.set_scene(env_idx, scene_params, thread_idx)
scene_params[4] *= 1000.
halfEdge = np.array([0.15, 0.8, 0.15])
center = np.array([0., 0., 0.])
quat = np.array([1., 0., 0., 0.])
pyflex.add_box(halfEdge, center, quat)
pyflex.add_box(halfEdge, center, quat)
n_particles = pyflex.get_n_particles()
n_shapes = pyflex.get_n_shapes()
positions = np.zeros((time_step, n_particles + n_shapes, 6), dtype=np.float32)
velocities = np.zeros((time_step, n_particles + n_shapes, 6), dtype=np.float32)
shape_quats = np.zeros((time_step, n_shapes, 4), dtype=np.float32)
for j in range(time_step_clip):
shape_states = calc_shape_states_RiceGrip(0, dt, shape_state_dim, gripper_config)
pyflex.set_shape_states(shape_states)
pyflex.step()
p = pyflex.get_positions().reshape(-1, 4)[:, :3]
clusters = []
st_time = time.time()
kmeans = MiniBatchKMeans(n_clusters=root_num[0][0], random_state=0).fit(p)
# print('Time on kmeans', time.time() - st_time)
clusters.append([[kmeans.labels_]])
# centers = kmeans.cluster_centers_
for j in range(time_step):
shape_states = calc_shape_states_RiceGrip(j * dt, dt, shape_state_dim, gripper_config)
pyflex.set_shape_states(shape_states)
positions[j, :n_particles, :3] = pyflex.get_rigidGlobalPositions().reshape(-1, 3)
positions[j, :n_particles, 3:] = pyflex.get_positions().reshape(-1, 4)[:, :3]
shape_states = pyflex.get_shape_states().reshape(-1, shape_state_dim)
for k in range(n_shapes):
positions[j, n_particles + k, :3] = shape_states[k, :3]
positions[j, n_particles + k, 3:] = shape_states[k, :3]
shape_quats[j, k] = shape_states[k, 6:10]
if j > 0:
velocities[j] = (positions[j] - positions[j - 1]) / dt
pyflex.step()
data = [positions[j], velocities[j], shape_quats[j], clusters, scene_params]
store_data(data_names, data, os.path.join(rollout_dir, str(j) + '.h5'))
else:
raise AssertionError("Unsupported env")
# change dtype for more accurate stat calculation
# only normalize positions and velocities
datas = [positions.astype(np.float64), velocities.astype(np.float64)]
# NOTE: stats is of length 2, for positions and velocities
for j in range(len(stats)):
stat = init_stat(stats[j].shape[0])
stat[:, 0] = np.mean(datas[j], axis=(0, 1))[:]
stat[:, 1] = np.std(datas[j], axis=(0, 1))[:]
stat[:, 2] = datas[j].shape[0] * datas[j].shape[1]
stats[j] = combine_stat(stats[j], stat)
pyflex.clean()
return stats
def gen_Cloth(info):
env, env_idx = info['env'], info['env_idx']
thread_idx, data_dir, data_names = info['thread_idx'], info['data_dir'], info['data_names']
n_rollout, time_step = info['n_rollout'], info['time_step']
dt, args, phase = info['dt'], info['args'], info['phase']
vis_width, vis_height = info['vis_width'], info['vis_height']
state_dim = args.state_dim
action_dim = args.action_dim
dt = 1. / 60.
np.random.seed(round(time.time() * 1000 + thread_idx) % 2 ** 32)
stats = [init_stat(state_dim), init_stat(action_dim)]
engine = ClothEngine(dt, state_dim, action_dim)
import pyflex
pyflex.init()
# bar = ProgressBar()
for i in range(n_rollout):
rollout_idx = thread_idx * n_rollout + i
rollout_dir = os.path.join(data_dir, str(rollout_idx))
os.system('mkdir -p ' + rollout_dir)
engine.init(pyflex)
scene_params = engine.scene_params
action = np.zeros(4)
states_all = np.zeros((time_step, engine.n_particles, state_dim))
actions_all = np.zeros((time_step, 1, action_dim))
# drop the cloth down
engine.set_action(action)
engine.step()
for j in range(time_step):
positions = pyflex.get_positions().reshape(-1, 4)[:, :3]
# sample the action
if j % 5 == 0:
ctrl_pts = rand_int(0, 8)
act_lim = 0.05
dx = rand_float(-act_lim, act_lim)
dz = rand_float(-act_lim, act_lim)
dy = 0.05
action = np.array([ctrl_pts, dx, dy, dz])
else:
action[2] = 0.
# store the rollout information
state = engine.get_state()
states_all[j] = state
tga_path = os.path.join(rollout_dir, '%d.tga' % j)
pyflex.render(capture=True, path=tga_path)
tga = Image.open(tga_path)
tga = np.array(tga)[:, 60:780, :3][:, :, ::-1]
tga = cv2.resize(tga, (vis_width, vis_height), interpolation=cv2.INTER_AREA)
os.system('rm ' + tga_path)
jpg_path = os.path.join(rollout_dir, 'fig_%d.jpg' % j)
cv2.imwrite(jpg_path, tga)
actions_all[j, 0] = action.copy()
engine.set_action(action)
engine.step()
datas = [states_all, actions_all, scene_params]
store_data(data_names, datas, rollout_dir + '.h5')
datas = [datas[j].astype(np.float64) for j in range(len(datas))]
for j in range(len(stats)):
stat = init_stat(stats[j].shape[0])
stat[:, 0] = np.mean(datas[j], axis=(0, 1))[:]
stat[:, 1] = np.std(datas[j], axis=(0, 1))[:]
stat[:, 2] = datas[j].shape[0]
stats[j] = combine_stat(stats[j], stat)
pyflex.clean()
return stats