def gripper(gripper, grasp, T_obj_world, color=(0.5, 0.5, 0.5), T_camera_world=None, point=None, point_color=None):
""" Plots a robotic gripper in a pose specified by a particular grasp object.
Parameters
----------
gripper : :obj:`dexnet.grasping.RobotGripper`
the gripper to plot
grasp : :obj:`dexnet.grasping.Grasp`
the grasp to plot the gripper performing
T_obj_world : :obj:`autolab_core.RigidTransform`
the pose of the object that the grasp is referencing in world frame
color : 3-tuple
color of the gripper mesh
"""
T_gripper_obj = grasp.gripper_pose(gripper)
T_gripper_world = T_obj_world * T_gripper_obj
T_mesh_world = T_gripper_world * gripper.T_mesh_gripper.inverse()
T_mesh_world = T_mesh_world.as_frames('obj', 'world')
Visualizer3D.mesh(gripper.mesh.trimesh, T_mesh_world=T_mesh_world, style='surface', color=color)
Visualizer3D.pose(T_gripper_world)
if T_camera_world is not None:
Visualizer3D.pose(T_camera_world)
if point is not None:
world_points = np.dot(T_obj_world.matrix, point.T).T
Visualizer3D.points(world_points[:, :3], scale=0.002, color=point_color)
def showpoint(surface_point,
T_obj_world=RigidTransform(from_frame='obj',
to_frame='world'),
color=(0.5, 0.5, 0),
scale=0.001):
""" Plots a grasp as an axis and center.
Parameters
----------
grasp : :obj:`dexnet.grasping.Grasp`
the grasp to plot
T_obj_world : :obj:`autolab_core.RigidTransform`
the pose of the object that the grasp is referencing in world frame
tube_radius : float
radius of the plotted grasp axis
endpoint_color : 3-tuple
color of the endpoints of the grasp axis
endpoint_scale : 3-tuple
scale of the plotted endpoints
grasp_axis_color : 3-tuple
color of the grasp axis
"""
surface_point = Point(surface_point, 'obj')
surface_point_tf = T_obj_world.apply(surface_point)
#center_tf = T_obj_world.apply(center)
Visualizer3D.points(surface_point_tf.data, color=color, scale=scale)
def main(args):
# set logging
logging.getLogger().setLevel(logging.INFO)
rospy.init_node("ensenso_reader", anonymous=True)
num_frames = 10
sensor = RgbdSensorFactory("ensenso", cfg={"frame": "ensenso"})
sensor.start()
total_time = 0
for i in range(num_frames):
if i > 0:
start_time = time.time()
_, depth_im, _ = sensor.frames()
if i > 0:
total_time += time.time() - start_time
print("Frame %d" % (i))
print("Avg FPS: %.5f" % (float(i) / total_time))
depth_im = sensor.median_depth_img(num_img=5)
point_cloud = sensor.ir_intrinsics.deproject(depth_im)
point_cloud.remove_zero_points()
sensor.stop()
vis2d.figure()
vis2d.imshow(depth_im)
vis2d.title("Ensenso - Raw")
vis2d.show()
vis3d.figure()
vis3d.points(point_cloud, random=True, subsample=10, scale=0.0025)
vis3d.show()
def grid_search(pc, indices, model, shadow, img_file):
length, width, height = shadow.extents
split_size = max(length, width)
pc_data, ind = get_pc_data(pc)
maxes = np.max(pc_data, axis=0)
mins = np.min(pc_data, axis=0)
bin_base = mins[2]
plane_normal = model[0:3]
scores = np.zeros((int(np.round((maxes[0] - mins[0]) / split_size)),
int(np.round((maxes[1] - mins[1]) / split_size))))
for i in range(int(np.round((maxes[0] - mins[0]) / split_size))):
x = mins[0] + i * split_size
for j in range(int(np.round((maxes[1] - mins[1]) / split_size))):
y = mins[1] + j * split_size
for sh in rotations(shadow, 8):
scores[i][j] = do_stuff(pc, indices, model, sh, img_file)
print("\nScores: \n" + str(scores))
best = best_cell(scores)
print("\nBest Cell: " + str(best) + ", with score = " +
str(scores[best[0]][best[1]]))
#-------
# Visualize best placement
vis3d.figure()
x = mins[0] + best[0] * split_size
y = mins[1] + best[1] * split_size
cell_indices = np.where((x < pc_data[:, 0])
& (pc_data[:, 0] < x + split_size)
& (y < pc_data[:, 1])
& (pc_data[:, 1] < y + split_size))[0]
points = pc_data[cell_indices]
rest = pc_data[np.setdiff1d(np.arange(len(pc_data)), cell_indices)]
vis3d.points(points, color=(0, 1, 1))
def show_points(points, color=(0,1,0), scale=0.005, frame_size=0.2, frame_radius=0.02):
vis.figure(bgcolor=(1,1,1), size=(500,500))
vis.points(np.array(points), color=color, scale=scale)
vis.plot3d(np.array(([0, 0, 0], [frame_size, 0, 0])).astype(np.float32), color=(1,0,0), tube_radius=frame_radius)
vis.plot3d(np.array(([0, 0, 0], [0, frame_size, 0])).astype(np.float32), color=(0,1,0), tube_radius=frame_radius)
vis.plot3d(np.array(([0, 0, 0], [0, 0, frame_size])).astype(np.float32), color=(0,0,1), tube_radius=frame_radius)
vis.show()
def fine_grid_search(pc, indices, model, shadow, splits):
length, width, height = shadow.extents
split_size = max(length, width)
pc_data, ind = get_pc_data(pc, indices)
maxes = np.max(pc_data, axis=0)
mins = np.min(pc_data, axis=0)
bin_base = mins[2]
plane_normal = model[0:3]
#splits = 3
step_size = split_size / splits
plane_data = get_plane_data(pc, indices)
plane_pc = PointCloud(plane_data.T, pc.frame)
plane_pc = cp.inverse().apply(plane_pc)
di = ci.project_to_image(plane_pc)
bi = di.to_binary()
bi = bi.inverse()
scene = Scene()
camera = VirtualCamera(ci, cp)
scene.camera = camera
shadow_obj = SceneObject(shadow)
scene.add_object('shadow', shadow_obj)
orig_tow = shadow_obj.T_obj_world
numx = (int(np.round((maxes[0]-mins[0])/split_size)) - 1) * splits + 1
numy = (int(np.round((maxes[1]-mins[1])/split_size)) - 1) * splits + 1
scores = np.zeros((numx, numy))
for i in range(numx):
x = mins[0] + i*step_size
for j in range(numy):
y = mins[1] + j*step_size
for tow in transforms(pc, pc_data, shadow, x, y, x+split_size, y+split_size, 8, orig_tow):
shadow_obj.T_obj_world = tow
scores[i][j] = under_shadow(scene, bi)
shadow_obj.T_obj_world = orig_tow
print("\nScores: \n" + str(scores))
best = best_cell(scores)
print("\nBest Cell: " + str(best) + ", with score = " + str(scores[best[0]][best[1]]))
#-------
# Visualize best placement
vis3d.figure()
x = mins[0] + best[0]*step_size
y = mins[1] + best[1]*step_size
cell_indices = np.where((x < pc_data[:,0]) & (pc_data[:,0] < x+split_size) & (y < pc_data[:,1]) & (pc_data[:,1] < y+split_size))[0]
points = pc_data[cell_indices]
rest = pc_data[np.setdiff1d(np.arange(len(pc_data)), cell_indices)]
vis3d.points(points, color=(0,1,1))
vis3d.points(rest, color=(1,0,1))
vis3d.show()
#--------
return best, scene
def visualize_scene(mesh, T_world_ar, T_ar_cam, T_cam_obj):
T_cam_cam = RigidTransform()
T_obj_cam = T_cam_obj.inverse()
T_cam_ar = T_ar_cam.inverse()
Twc = np.matmul(T_world_ar.matrix, T_ar_cam.matrix)
T_world_cam = RigidTransform(Twc[:3, :3], Twc[:3, 3])
T_cam_world = T_world_cam.inverse()
o = np.array([0., 0., 0.])
centroid_cam = apply_transform(o, T_cam_obj)
tag_cam = apply_transform(o, T_cam_ar)
robot_cam = apply_transform(o, T_cam_world)
# camera
vis3d.points(o, color=(0, 1, 0), scale=0.01)
vis3d.pose(T_cam_cam, alpha=0.05, tube_radius=0.005, center_scale=0.002)
# object
vis3d.mesh(mesh)
vis3d.points(centroid_cam, color=(0, 0, 0), scale=0.01)
vis3d.pose(T_cam_obj, alpha=0.05, tube_radius=0.005, center_scale=0.002)
# AR tag
vis3d.points(tag_cam, color=(1, 0, 1), scale=0.01)
vis3d.pose(T_cam_ar, alpha=0.05, tube_radius=0.005, center_scale=0.002)
vis3d.table(T_cam_ar, dim=0.074)
# robot
vis3d.points(robot_cam, color=(1, 0, 0), scale=0.01)
vis3d.pose(T_cam_world, alpha=0.05, tube_radius=0.005, center_scale=0.002)
vis3d.show()
def visualize_normals(mesh, vertices, normals):
scale = 0.01
normals_scaled = normals * scale
vis3d.pose(RigidTransform(),
alpha=0.01,
tube_radius=0.001,
center_scale=0.002)
vis3d.mesh(mesh, style='wireframe')
vis3d.points(mesh.centroid, color=(0, 0, 0), scale=0.003)
vis3d.points(vertices, color=(1, 0, 0), scale=0.001)
for v, n in zip(vertices, normals_scaled):
vis3d.plot3d([v, v + n], color=(1, 0, 0), tube_radius=0.0005)
vis3d.show()
def main():
logging.getLogger().setLevel(logging.INFO)
# parse args
parser = argparse.ArgumentParser(description='Register a webcam to the Photoneo PhoXi')
parser.add_argument('--config_filename', type=str, default='cfg/tools/colorize_phoxi.yaml', help='filename of a YAML configuration for registration')
args = parser.parse_args()
config_filename = args.config_filename
config = YamlConfig(config_filename)
sensor_data = config['sensors']
phoxi_config = sensor_data['phoxi']
phoxi_config['frame'] = 'phoxi'
# Initialize ROS node
rospy.init_node('colorize_phoxi', anonymous=True)
logging.getLogger().addHandler(rl.RosStreamHandler())
# Get PhoXi sensor set up
phoxi = RgbdSensorFactory.sensor(phoxi_config['type'], phoxi_config)
phoxi.start()
# Capture PhoXi and webcam images
phoxi_color_im, phoxi_depth_im, _ = phoxi.frames()
# vis2d.figure()
# vis2d.subplot(121)
# vis2d.imshow(phoxi_color_im)
# vis2d.subplot(122)
# vis2d.imshow(phoxi_depth_im)
# vis2d.show()
phoxi_pc = phoxi.ir_intrinsics.deproject(phoxi_depth_im)
colors = phoxi_color_im.data.reshape((phoxi_color_im.shape[0] * phoxi_color_im.shape[1], phoxi_color_im.shape[2])) / 255.0
vis3d.figure()
vis3d.points(phoxi_pc.data.T[::3], color=colors[::3], scale=0.001)
vis3d.show()
# Export to PLY file
vertices = phoxi.ir_intrinsics.deproject(phoxi_depth_im).data.T
colors = phoxi_color_im.data.reshape(phoxi_color_im.data.shape[0] * phoxi_color_im.data.shape[1], phoxi_color_im.data.shape[2])
f = open('pcloud.ply', 'w')
f.write('ply\nformat ascii 1.0\nelement vertex {}\nproperty float x\nproperty float y\nproperty float z\nproperty uchar red\n'.format(len(vertices)) +
'property uchar green\nproperty uchar blue\nend_header\n')
for v, c in zip(vertices,colors):
f.write('{} {} {} {} {} {}\n'.format(v[0], v[1], v[2], c[0], c[1], c[2]))
f.close()
def grasp(grasp,
T_obj_world=RigidTransform(from_frame='obj', to_frame='world'),
tube_radius=0.0002,
endpoint_color=(0, 1, 0),
endpoint_scale=0.0005,
grasp_axis_color=(0, 1, 0)):
""" Plots a grasp as an axis and center.
Parameters
----------
grasp : :obj:`dexnet.grasping.Grasp`
the grasp to plot
T_obj_world : :obj:`autolab_core.RigidTransform`
the pose of the object that the grasp is referencing in world frame
tube_radius : float
radius of the plotted grasp axis
endpoint_color : 3-tuple
color of the endpoints of the grasp axis
endpoint_scale : 3-tuple
scale of the plotted endpoints
grasp_axis_color : 3-tuple
color of the grasp axis
"""
g1, g2 = grasp.endpoints
center = grasp.center
g1 = Point(g1, 'obj')
g2 = Point(g2, 'obj')
center = Point(center, 'obj')
g1_tf = T_obj_world.apply(g1)
g2_tf = T_obj_world.apply(g2)
center_tf = T_obj_world.apply(center)
grasp_axis_tf = np.array([g1_tf.data, g2_tf.data])
Visualizer3D.points(g1_tf.data,
color=endpoint_color,
scale=endpoint_scale)
Visualizer3D.points(g2_tf.data,
color=endpoint_color,
scale=endpoint_scale)
Visualizer3D.plot3d(grasp_axis_tf,
color=grasp_axis_color,
tube_radius=tube_radius)
Visualizer3D.pose(grasp.T_grasp_obj,
alpha=endpoint_scale * 10,
tube_radius=tube_radius,
center_scale=endpoint_scale)
def grid_search(pc, indices, model, shadow, img_file):
length, width, height = shadow.extents
split_size = max(length, width)
pc_data, ind = get_pc_data(pc, indices)
maxes = np.max(pc_data, axis=0)
mins = np.min(pc_data, axis=0)
bin_base = mins[2]
plane_normal = model[0:3]
scores = np.zeros((int(np.round((maxes[0] - mins[0]) / split_size)),
int(np.round((maxes[1] - mins[1]) / split_size))))
for i in range(int(np.round((maxes[0] - mins[0]) / split_size))):
x = mins[0] + i * split_size
for j in range(int(np.round((maxes[1] - mins[1]) / split_size))):
y = mins[1] + j * split_size
#binarized_overlap_image(pc, x, y, x+split_size, y+split_size, shadow, plane_normal, indices, model)
for sh in rotations(shadow, 8):
#overlap_size = binarized_overlap_image(pc, x, y, x+split_size, y+split_size, sh, plane_normal, indices, model)
#scores[i][j] = -1*overlap_size
scene = Scene()
camera = VirtualCamera(ci, cp)
scene.camera = camera
scores[i][j] = under_shadow(pc, pc_data, indices, model, sh, x,
x + split_size, y, y + split_size,
scene)
print("\nScores: \n" + str(scores))
best = best_cell(scores)
print("\nBest Cell: " + str(best) + ", with score = " +
str(scores[best[0]][best[1]]))
#-------
# Visualize best placement
vis3d.figure()
x = mins[0] + best[0] * split_size
y = mins[1] + best[1] * split_size
cell_indices = np.where((x < pc_data[:, 0])
& (pc_data[:, 0] < x + split_size)
& (y < pc_data[:, 1])
& (pc_data[:, 1] < y + split_size))[0]
points = pc_data[cell_indices]
rest = pc_data[np.setdiff1d(np.arange(len(pc_data)), cell_indices)]
vis3d.points(points, color=(0, 1, 1))
vis3d.points(rest, color=(1, 0, 1))
vis3d.show()
def vis(self, mesh, grasp_vertices, grasp_qualities, grasp_normals):
"""
Pass in any grasp and its associated grasp quality. this function will plot
each grasp on the object and plot the grasps as a bar between the points, with
colored dots on the line endpoints representing the grasp quality associated
with each grasp
Parameters
----------
mesh : :obj:`Trimesh`
grasp_vertices : mx2x3 :obj:`numpy.ndarray`
m grasps. Each grasp containts two contact points. Each contact point
is a 3 dimensional vector, hence the shape mx2x3
grasp_qualities : mx' :obj:`numpy.ndarray`
vector of grasp qualities for each grasp
"""
vis3d.mesh(mesh)
middle_of_part = np.mean(np.mean(grasp_vertices, axis=1), axis=0)
print(middle_of_part)
vis3d.points(middle_of_part, scale=0.003)
dirs = normalize(grasp_vertices[:, 0] - grasp_vertices[:, 1], axis=1)
midpoints = (grasp_vertices[:, 0] + grasp_vertices[:, 1]) / 2
grasp_endpoints = np.zeros(grasp_vertices.shape)
grasp_endpoints[:, 0] = midpoints + dirs * MAX_HAND_DISTANCE / 2
grasp_endpoints[:, 1] = midpoints - dirs * MAX_HAND_DISTANCE / 2
n0 = np.zeros(grasp_endpoints.shape)
n1 = np.zeros(grasp_endpoints.shape)
normal_scale = 0.01
n0[:, 0] = grasp_vertices[:, 0]
n0[:, 1] = grasp_vertices[:, 0] + normal_scale * grasp_normals[:, 0]
n1[:, 0] = grasp_vertices[:, 1]
n1[:, 1] = grasp_vertices[:, 1] + normal_scale * grasp_normals[:, 1]
for grasp, quality, normal0, normal1 in zip(grasp_endpoints,
grasp_qualities, n0, n1):
color = [min(1, 2 * (1 - quality)), min(1, 2 * quality), 0, 1]
vis3d.plot3d(grasp, color=color, tube_radius=.001)
vis3d.plot3d(normal0, color=(0, 0, 0), tube_radius=.002)
vis3d.plot3d(normal1, color=(0, 0, 0), tube_radius=.002)
vis3d.show()
def visualize_grasps(mesh, vertices, metrics):
vis3d.pose(RigidTransform(),
alpha=0.01,
tube_radius=0.001,
center_scale=0.002)
vis3d.mesh(mesh, style='wireframe')
vis3d.points(mesh.centroid, color=(0, 0, 0), scale=0.003)
min_score = float(np.min(metrics))
max_score = float(np.max(metrics))
metrics_normalized = (metrics.astype(float) - min_score) / (max_score -
min_score)
for v, m in zip(vertices, metrics_normalized):
vis3d.points(v, color=(1 - m, m, 0), scale=0.001)
vis3d.plot3d(v, color=(1 - m, m, 0), tube_radius=0.0003)
vis3d.show()
def visualize_test():
I = RigidTransform()
g_ab = RigidTransform()
g_ab.translation = np.array([0.05, 0, 0])
g_ab.rotation = np.array([[0, -1, 0], [1, 0, 0], [0, 0, 1]])
q_a = np.array([0., 0., 0.])
p_b = np.array([0., 0., 0.])
p_a = apply_transform(p_b, g_ab)
print('g_ab = \n{}'.format(g_ab.matrix))
vis3d.pose(I, alpha=0.01, tube_radius=0.001, center_scale=0.001)
vis3d.points(q_a, color=(1, 0, 0), scale=0.005)
vis3d.pose(g_ab, alpha=0.01, tube_radius=0.001, center_scale=0.001)
vis3d.points(p_a, color=(0, 1, 0), scale=0.005)
vis3d.show()
def main():
start_time = time.time()
img_file = '/nfs/diskstation/projects/dex-net/placing/datasets/real/sample_ims_05_22/depth_ims_numpy/image_000001.npy'
ci_file = '/nfs/diskstation/projects/dex-net/placing/datasets/real/sample_ims_05_22/camera_intrinsics.intr'
mesh_file = 'demon_helmet.obj'
indices, model, image, pc = largest_planar_surface(img_file, ci_file)
mesh, best_pose, rt = find_stable_poses(mesh_file)
shadow = find_shadow(mesh, best_pose, model[0:3])
vis3d.figure()
vis3d.points(pc, color=(1, 0, 0))
vis3d.mesh(shadow, rt)
vis3d.show()
scores, split_size = score_cells(pc, indices, model, shadow, ci_file)
ind = best_cell(scores)
# print("Scores: \n" + str(scores))
# print("\nBest cell = " + str(ind))
print("--- %s seconds ---" % (time.time() - start_time))
def visualize_plan(mesh, T_world_obj, T_world_grasp):
# visualize the plan in the world frame
mesh.apply_transform(T_world_obj.matrix)
o = np.array([0., 0., 0.])
obj = apply_transform(o, T_world_obj)
grasp = apply_transform(o, T_world_grasp)
# base frame
vis3d.points(o, color=(1, 0, 0), scale=0.005)
vis3d.pose(RigidTransform(),
alpha=0.03,
tube_radius=0.002,
center_scale=0.001)
# object
vis3d.mesh(mesh)
vis3d.points(obj, color=(0, 0, 0), scale=0.005)
vis3d.pose(T_world_obj, alpha=0.03, tube_radius=0.002, center_scale=0.001)
# grasp
vis3d.points(grasp, color=(0, 1, 1), scale=0.005)
vis3d.pose(T_world_grasp,
alpha=0.03,
tube_radius=0.002,
center_scale=0.001)
vis3d.show()
def main(args):
# set logging
logging.getLogger().setLevel(logging.INFO)
rospy.init_node('ensenso_reader', anonymous=True)
num_frames = 10
#sensor = PhoXiSensor(frame='phoxi',
# size='small')
sensor = EnsensoSensor(frame='ensenso')
sensor.start()
total_time = 0
for i in range(num_frames):
if i > 0:
start_time = time.time()
_, depth_im, _ = sensor.frames()
if i > 0:
total_time += time.time() - start_time
print('Frame %d' % (i))
print('Avg FPS: %.5f' % (float(i) / total_time))
depth_im = sensor.median_depth_img(num_img=5)
point_cloud = sensor.ir_intrinsics.deproject(depth_im)
point_cloud.remove_zero_points()
sensor.stop()
vis2d.figure()
vis2d.imshow(depth_im)
vis2d.title('PhoXi - Raw')
vis2d.show()
vis3d.figure()
vis3d.points(point_cloud, random=True, subsample=10, scale=0.0025)
vis3d.show()
def visualize_gripper(mesh, T_obj_grasp, vertices):
o = np.array([0., 0., 0.])
grasp = apply_transform(o, T_obj_grasp)
# object
vis3d.mesh(mesh)
vis3d.points(o, color=(0, 0, 0), scale=0.002)
vis3d.pose(RigidTransform(),
alpha=0.01,
tube_radius=0.001,
center_scale=0.001)
# gripper
vis3d.points(grasp, color=(1, 0, 0), scale=0.002)
vis3d.points(vertices, color=(1, 0, 0), scale=0.002)
vis3d.pose(T_obj_grasp, alpha=0.01, tube_radius=0.001, center_scale=0.001)
vis3d.show()
def vis(self, mesh, grasp_vertices, grasp_qualities, top_n_grasp_vertices,
approach_directions, rs, TG):
"""
Pass in any grasp and its associated grasp quality. this function will plot
each grasp on the object and plot the grasps as a bar between the points, with
colored dots on the line endpoints representing the grasp quality associated
with each grasp
Parameters
----------
mesh : :obj:`Trimesh`
grasp_vertices : mx2x3 :obj:`numpy.ndarray`
m grasps. Each grasp containts two contact points. Each contact point
is a 3 dimensional vector, hence the shape mx2x3
grasp_qualities : mx' :obj:`numpy.ndarray`
vector of grasp qualities for each grasp
"""
vis3d.mesh(mesh)
dirs = normalize(grasp_vertices[:, 0] - grasp_vertices[:, 1], axis=1)
midpoints = (grasp_vertices[:, 0] + grasp_vertices[:, 1]) / 2
grasp_endpoints = np.zeros(grasp_vertices.shape)
grasp_vertices[:, 0] = midpoints + dirs * MAX_HAND_DISTANCE / 2
grasp_vertices[:, 1] = midpoints - dirs * MAX_HAND_DISTANCE / 2
for i, (grasp,
quality) in enumerate(zip(grasp_vertices, grasp_qualities)):
color = [min(1, 2 * (1 - quality)), min(1, 2 * quality), 0, 1]
vis3d.plot3d(grasp, color=color, tube_radius=.001)
blue = [0, 0, 255]
light_blue = [50, 50, 200]
for i, (grasp, approach_direction) in enumerate(
zip(top_n_grasp_vertices, approach_directions)):
midpoint = np.mean(grasp, axis=0)
approach_direction = np.asarray(
[midpoint, midpoint - 0.1 * approach_direction])
if (i == 0):
vis3d.plot3d(approach_direction, color=blue, tube_radius=.005)
vis3d.points(grasp, color=blue, scale=.005)
else:
vis3d.plot3d(approach_direction,
color=light_blue,
tube_radius=.001)
vis3d.points(grasp, color=light_blue, scale=.001)
midpoint = np.mean(top_n_grasp_vertices[0], axis=0)
x_purp = [204, 0, 204]
x_axis = np.asarray([midpoint, midpoint + 0.1 * rs[:, 0]])
y_cyan = [0, 204, 204]
y_axis = np.asarray([midpoint, midpoint + 0.1 * rs[:, 1]])
z_black = [0, 0, 0]
z_axis = np.asarray([midpoint, midpoint + 0.3 * rs[:, 2]])
vis3d.plot3d(x_axis, color=x_purp, tube_radius=.005)
vis3d.plot3d(y_axis, color=y_cyan, tube_radius=.005)
vis3d.plot3d(z_axis, color=z_black, tube_radius=.005)
origin = np.asarray([0, 0, 0])
x_axis = np.asarray([origin, 0.2 * np.array([1, 0, 0])])
y_axis = np.asarray([origin, 0.2 * np.array([0, 1, 0])])
z_axis = np.asarray([origin, 0.2 * np.array([0, 0, 1])])
vis3d.plot3d(x_axis, color=x_purp, tube_radius=.005)
vis3d.plot3d(y_axis, color=y_cyan, tube_radius=.005)
vis3d.plot3d(z_axis, color=z_black, tube_radius=.005)
# eucl_orien = np.asarray(tfs.euler_from_quaternion(TG.quaternion))
intermediate_pos = TG.position - np.reshape(
np.matmul(TG.rotation, np.array([[0], [0], [0.2]])), (1, 3))
print(intermediate_pos, np.shape(intermediate_pos))
red = [255, 0, 0]
vis3d.points(intermediate_pos, color=red, scale=.01)
vis3d.points(TG.position, color=red, scale=.01)
vis3d.show()
output_filename = os.path.join(
output_path, '{0}_to_world.tf'.format(T_world_obj.from_frame))
print T_world_obj
T_world_obj.save(output_filename)
if config['vis'] and VIS_SUPPORTED:
_, depth_im, _ = sensor.frames()
pc_cam = ir_intrinsics.deproject(depth_im)
pc_world = T_world_cam * pc_cam
mesh_file = ObjFile(
os.path.join(object_path, '{0}.obj'.format(args.object_name)))
mesh = mesh_file.read()
vis.figure(bgcolor=(0.7, 0.7, 0.7))
vis.mesh(mesh, T_world_obj.as_frames('obj', 'world'), style='surface')
vis.pose(T_world_obj, alpha=0.04, tube_radius=0.002, center_scale=0.01)
vis.pose(RigidTransform(from_frame='origin'),
alpha=0.04,
tube_radius=0.002,
center_scale=0.01)
vis.pose(T_world_cam, alpha=0.04, tube_radius=0.002, center_scale=0.01)
vis.pose(T_world_cam * T_cb_cam.inverse(),
alpha=0.04,
tube_radius=0.002,
center_scale=0.01)
vis.points(pc_world, subsample=20)
vis.show()
sensor.stop()
logging.info('Finding closest orthogonal grasp')
T_grasp_world = get_closest_grasp_pose(T_tag_world, T_ready_world)
T_lift = RigidTransform(translation=[0, 0, 0.2],
from_frame=T_ready_world.to_frame,
to_frame=T_ready_world.to_frame)
T_lift_world = T_lift * T_grasp_world
logging.info('Visualizing poses')
_, depth_im, _ = sensor.frames()
points_world = T_camera_world * intr.deproject(depth_im)
if cfg['vis_detect']:
vis3d.figure()
vis3d.pose(RigidTransform())
vis3d.points(subsample(points_world.data.T, 0.01),
color=(0, 1, 0),
scale=0.002)
vis3d.pose(T_ready_world, length=0.05)
vis3d.pose(T_camera_world, length=0.1)
vis3d.pose(T_tag_world)
vis3d.pose(T_grasp_world)
vis3d.pose(T_lift_world)
vis3d.show()
#const_rotation=np.array([[1,0,0],[0,-1,0],[0,0,-1]])
#test = RigidTransform(rotation=const_rotation,translation=T_tag_world.translation, from_frame='franka_tool', to_frame='world')
#import pdb; pdb.set_trace()
rotation = T_tag_world.rotation
rotation[0:2, :] = -1 * rotation[0:2, :]
def compute_approach_direction(self, mesh, grasp_vertices, grasp_quality,
grasp_normals):
## initalizing stuff ##
visualize = True
nb_directions_to_test = 6
normal_scale = 0.01
plane_normal = normalize(grasp_vertices[0] - grasp_vertices[1])
midpoint = (grasp_vertices[0] + grasp_vertices[1]) / 2
## generating a certain number of approach directions ##
theta = np.pi / nb_directions_to_test
rot_mat = rotation_3d(-plane_normal, theta)
horizontal_direction = normalize(
np.cross(plane_normal, np.array([0, 0, 1])))
directions_to_test = [horizontal_direction] #these are vectors
approach_directions = [
np.array(
[midpoint, midpoint + horizontal_direction * normal_scale])
] #these are two points for visualization
for i in range(nb_directions_to_test - 1):
directions_to_test.append(
normalize(np.matmul(rot_mat, directions_to_test[-1])))
approach_directions.append(
np.array([
midpoint, midpoint + directions_to_test[-1] * normal_scale
]))
## computing the palm position for each approach direction ##
palm_positions = []
for i in range(nb_directions_to_test):
palm_positions.append(midpoint +
finger_length * directions_to_test[i])
if visualize:
## plotting the whole mesh ##
vis3d.mesh(mesh, style='wireframe')
## computing and plotting midpoint and gripper position ##
dirs = (grasp_vertices[0] - grasp_vertices[1]
) / np.linalg.norm(grasp_vertices[0] - grasp_vertices[1])
grasp_endpoints = np.zeros(grasp_vertices.shape)
grasp_endpoints[0] = midpoint + dirs * MAX_HAND_DISTANCE / 2
grasp_endpoints[1] = midpoint - dirs * MAX_HAND_DISTANCE / 2
color = [
min(1, 2 * (1 - grasp_quality)),
min(1, 2 * grasp_quality), 0, 1
]
vis3d.plot3d(grasp_endpoints, color=color, tube_radius=.001)
vis3d.points(midpoint, scale=0.003)
## computing and plotting normals at contact points ##
n0 = np.zeros(grasp_endpoints.shape)
n1 = np.zeros(grasp_endpoints.shape)
n0[0] = grasp_vertices[0]
n0[1] = grasp_vertices[0] + normal_scale * grasp_normals[0]
n1[0] = grasp_vertices[1]
n1[1] = grasp_vertices[1] + normal_scale * grasp_normals[1]
vis3d.plot3d(n0, color=(0, 0, 0), tube_radius=.002)
vis3d.plot3d(n1, color=(0, 0, 0), tube_radius=.002)
## plotting normals the palm positions for each potential approach direction ##
for i in range(nb_directions_to_test):
vis3d.points(palm_positions[i], scale=0.003)
vis3d.show()
directions_to_test = [
directions_to_test[3], directions_to_test[2],
directions_to_test[4], directions_to_test[1],
directions_to_test[5], directions_to_test[0]
]
palm_positions = [
palm_positions[3], palm_positions[2], palm_positions[4],
palm_positions[1], palm_positions[5], palm_positions[0]
]
## checking if some approach direction is valid ##
for i in range(nb_directions_to_test):
if len(
trimesh.intersections.mesh_plane(mesh,
directions_to_test[i],
palm_positions[i])) == 0:
# it means the palm won't bump with part
return directions_to_test[i]
# it means all approach directions will bump with part
return -1
KSIZE = 9
if __name__ == '__main__':
depth_im_filename = sys.argv[1]
camera_intr_filename = sys.argv[2]
camera_intr = CameraIntrinsics.load(camera_intr_filename)
depth_im = DepthImage.open(depth_im_filename, frame=camera_intr.frame)
depth_im = depth_im.inpaint()
point_cloud_im = camera_intr.deproject_to_image(depth_im)
normal_cloud_im = point_cloud_im.normal_cloud_im(ksize=KSIZE)
vis3d.figure()
vis3d.points(point_cloud_im.to_point_cloud(), scale=0.0025)
alpha = 0.025
subsample = 20
for i in range(0, point_cloud_im.height, subsample):
for j in range(0, point_cloud_im.width, subsample):
p = point_cloud_im[i, j]
n = normal_cloud_im[i, j]
n2 = normal_cloud_im_s[i, j]
if np.linalg.norm(n) > 0:
points = np.array([p, p + alpha * n])
vis3d.plot3d(points, tube_radius=0.001, color=(1, 0, 0))
points = np.array([p, p + alpha * n2])
vis3d.plot3d(points, tube_radius=0.001, color=(1, 0, 1))
# vis
if config['vis_points']:
_, depth_im, _ = sensor.frames()
points_world = T_camera_world * ir_intrinsics.deproject(depth_im)
true_robot_points_world = PointCloud(np.array([T.translation for T in robot_poses]).T,
frame=ir_intrinsics.frame)
est_robot_points_world = T_camera_world * PointCloud(np.array(robot_points_camera).T,
frame=ir_intrinsics.frame)
mean_est_robot_point = np.mean(est_robot_points_world.data, axis=1).reshape(3,1)
est_robot_points_world._data = est_robot_points_world._data - mean_est_robot_point + mean_true_robot_point
fixed_robot_points_world = T_corrected_cb_world * est_robot_points_world
mean_fixed_robot_point = np.mean(fixed_robot_points_world.data, axis=1).reshape(3,1)
fixed_robot_points_world._data = fixed_robot_points_world._data - mean_fixed_robot_point + mean_true_robot_point
vis3d.figure()
vis3d.points(points_world, color=(0,1,0), subsample=10, random=True, scale=0.001)
vis3d.points(true_robot_points_world, color=(0,0,1), scale=0.001)
vis3d.points(fixed_robot_points_world, color=(1,1,0), scale=0.001)
vis3d.points(est_robot_points_world, color=(1,0,0), scale=0.001)
vis3d.pose(T_camera_world)
vis3d.show()
# save tranformation arrays based on setup
output_dir = os.path.join(config['calib_dir'], sensor_frame)
if not os.path.exists(output_dir):
os.makedirs(output_dir)
pose_filename = os.path.join(output_dir, '%s_to_world.tf' %(sensor_frame))
T_camera_world.save(pose_filename)
intr_filename = os.path.join(output_dir, '%s.intr' %(sensor_frame))
ir_intrinsics.save(intr_filename)
f = open(os.path.join(output_dir, 'corners_cb_%s.npy' %(sensor_frame)), 'w')
# filter high
high_indices = np.where(point_cloud_world.data[2, :] > max_height)[0]
point_cloud_filtered.data[2, high_indices] = max_height
# re-project and update depth im
#depth_im_filtered = camera_intr.project_to_image(T_camera_world.inverse() * point_cloud_filtered)
logging.info('Clipping took %.3f sec' % (time.time() - clip_start))
# vis
focal_point = np.mean(point_cloud_filtered.data, axis=1)
if vis_clipping:
vis3d.figure(camera_pose=T_camera_world.as_frames('camera', 'world'),
focal_point=focal_point)
vis3d.points(point_cloud_world,
scale=0.001,
color=(1, 0, 0),
subsample=10)
vis3d.points(point_cloud_filtered,
scale=0.001,
color=(0, 0, 1),
subsample=10)
vis3d.show()
pcl_start = time.time()
# subsample point cloud
#rate = int(1.0 / rescale_factor)**2
#point_cloud_filtered = point_cloud_filtered.subsample(rate, random=False)
box = Box(np.array([0.2, -0.24, min_height]),
np.array([0.56, 0.21, max_height]),
frame='world')
di = DepthImage(image, frame=ci.frame)
pc = ci.deproject(di)
## Visualize the depth image
#vis2d.figure()
#vis2d.imshow(di)
#vis2d.show()
# Make and display a PCL type point cloud from the image
p = pcl.PointCloud(pc.data.T.astype(np.float32))
# Make a segmenter and segment the point cloud.
seg = p.make_segmenter()
seg.set_model_type(pcl.SACMODEL_PARALLEL_PLANE)
seg.set_method_type(pcl.SAC_RANSAC)
seg.set_distance_threshold(0.005)
indices, model = seg.segment()
print(model)
#pdb.set_trace()
vis3d.figure()
pc_plane = pc.data.T[indices]
pc_plane = pc_plane[np.where(pc_plane[::, 1] < 0.16)]
maxes = np.max(pc_plane, axis=0)
mins = np.min(pc_plane, axis=0)
print('maxes are :', maxes, '\nmins are : ', mins)
vis3d.points(pc_plane, color=(1, 0, 0))
vis3d.show()
# from_frame='obj', to_frame='table')
T_obj_table = RigidTransform(
rotation=[-0.1335021, 0.87671711, 0.41438141, 0.20452958],
from_frame='obj',
to_frame='table')
stable_pose = mesh.resting_pose(T_obj_table)
#print stable_pose.r
table_dim = 0.3
T_obj_table_plot = mesh.get_T_surface_obj(T_obj_table)
T_obj_table_plot.translation[0] += 0.1
vis.figure()
vis.mesh(mesh, T_obj_table_plot, color=(1, 0, 0), style='wireframe')
vis.points(Point(mesh.center_of_mass, 'obj'),
T_obj_table_plot,
color=(1, 0, 1),
scale=0.01)
vis.pose(T_obj_table_plot, alpha=0.1)
vis.mesh_stable_pose(mesh,
stable_pose,
dim=table_dim,
color=(0, 1, 0),
style='surface')
vis.pose(stable_pose.T_obj_table, alpha=0.1)
vis.show()
exit(0)
# compute stable poses
vis.figure()
vis.mesh(mesh, color=(1, 1, 0), style='surface')
vis.mesh(mesh.convex_hull(), color=(1, 0, 0))
vis.figure(size=(10, 10))
num_plot = 3
vis.subplot(1, num_plot, 1)
vis.imshow(depth_im)
vis.subplot(1, num_plot, 2)
vis.imshow(segmask)
vis.subplot(1, num_plot, 3)
vis.imshow(obj_segmask)
vis.show()
from visualization import Visualizer3D as vis3d
point_cloud = camera_intr.deproject(depth_im)
vis3d.figure()
vis3d.points(point_cloud,
subsample=3,
random=True,
color=(0, 0, 1),
scale=0.001)
vis3d.pose(RigidTransform())
vis3d.pose(T_camera_world.inverse())
vis3d.show()
# Create state.
rgbd_im = RgbdImage.from_color_and_depth(color_im, depth_im)
state = RgbdImageState(rgbd_im, camera_intr, segmask=segmask)
# Init policy.
policy_type = "cem"
if "type" in policy_config:
policy_type = policy_config["type"]
if policy_type == "ranking":
def visualize_vertices(mesh, vertices):
vis3d.mesh(mesh, style='wireframe')
vis3d.points(mesh.centroid, color=(0, 0, 0), scale=0.003)
vis3d.points(vertices, color=(1, 0, 0), scale=0.001)
vis3d.show()
vertices = mesh.vertices
triangles = mesh.triangles
normals = mesh.normals
print 'Num vertices:', len(vertices)
print 'Num triangles:', len(triangles)
print 'Num normals:', len(normals)
# 1. Generate candidate pairs of contact points
# 2. Check for force closure
# 3. Convert each grasp to a hand pose
contact1 = vertices[500]
contact2 = vertices[2000]
T_obj_gripper = contacts_to_baxter_hand_pose(contact1, contact2)
print 'Translation', T_obj_gripper.translation
print 'Rotation', T_obj_gripper.quaternion
pose_msg = T_obj_gripper.pose_msg
# 3d visualization to help debug
vis.figure()
vis.mesh(mesh)
vis.points(Point(contact1, frame='test'))
vis.points(Point(contact2, frame='test'))
vis.pose(T_obj_gripper, alpha=0.05)
vis.show()
# 4. Execute on the actual robot
