def figure_0():
action = policy.action(env.state)
env.render_3d_scene()
bottom_points = policy.toppling_model.bottom_points
vis3d.plot3d(bottom_points[:2], color=[0, 0, 0], tube_radius=.001)
mesh = env.state.mesh.copy().apply_transform(env.state.T_obj_world.matrix)
mesh.fix_normals()
direction = normalize([0, -.04, 0])
origin = mesh.center_mass + np.array([0, .04, .09])
intersect, _, face_ind = mesh.ray.intersects_location([origin],
[direction],
multiple_hits=False)
normal = mesh.face_normals[face_ind[0]]
start_point = intersect[0] + .06 * normal
end_point = intersect[0]
shaft_points = [start_point, end_point]
h1 = np.array([[1, 0, 0], [0, 0.7071, -0.7071], [0, 0.7071,
0.7071]]).dot(-normal)
h2 = np.array([[1, 0, 0], [0, 0.7071, 0.7071], [0, -0.7071,
0.7071]]).dot(-normal)
head_points = [end_point - 0.02 * h2, end_point, end_point - 0.02 * h1]
vis3d.plot3d(shaft_points, color=[1, 0, 0], tube_radius=.002)
vis3d.plot3d(head_points, color=[1, 0, 0], tube_radius=.002)
vis3d.points(Point(end_point), scale=.004, color=[0, 0, 0])
hand_pose = RigidTransform(rotation=policy.get_hand_pose(
start_point, end_point)[0].rotation,
translation=start_point,
from_frame='grasp',
to_frame='world')
orig_pose = env.state.obj.T_obj_world.copy()
env.state.obj.T_obj_world = policy.toppling_model.final_poses[1]
action = policy.grasping_policy.action(env.state)
env.state.obj.T_obj_world = orig_pose
gripper = env.gripper(action)
#vis3d.mesh(gripper.mesh, hand_pose * gripper.T_mesh_grasp, color=light_blue)
# mesh = trimesh.load('~/Downloads/chris.stl')
rot = np.array([[0, -1, 0], [1, 0, 0],
[0, 0,
1]]).dot(np.array([[0, 0, 1], [0, 1, 0], [-1, 0, 0]]))
T = RigidTransform(rotation=rot,
translation=np.array([0, -.09, .1]),
to_frame='mesh',
from_frame='mesh')
# vis3d.mesh(mesh, hand_pose * gripper.T_mesh_grasp * T, color=light_blue)
vis3d.mesh(gripper.mesh,
hand_pose * gripper.T_mesh_grasp * T,
color=light_blue)
vis3d.show(starting_camera_pose=CAMERA_POSE)
env.state.obj.T_obj_world = policy.toppling_model.final_poses[1]
action = policy.grasping_policy.action(env.state)
print 'q:', action.q_value
env.render_3d_scene()
vis3d.gripper(env.gripper(action),
action.grasp(env.gripper(action)),
color=light_blue)
vis3d.show(starting_camera_pose=CAMERA_POSE)
def save_samples(self, sample, grasps, T_obj_stp, collision_checker, grasp_metrics, stable_pose):
T_stp_camera = sample.camera.object_to_camera_pose
self.T_obj_camera = T_stp_camera * T_obj_stp.as_frames('obj', T_stp_camera.from_frame)
aligned_grasps = self.align_grasps_with_camera(grasps)
depth_im_table = sample.renders[RenderMode.DEPTH_SCENE].image
camera_pose = self.get_camera_pose(sample)
self.tensor_datapoint['camera_poses'] = camera_pose
shifted_camera_intr = sample.camera.camera_intr
for cnt, aligned_grasp in enumerate(aligned_grasps):
collision_free = self.is_grasp_collision_free(aligned_grasp, collision_checker)
# Project grasp coordinates in image
grasp_2d = aligned_grasp.project_camera(self.T_obj_camera, shifted_camera_intr)
grasp_point = aligned_grasp.close_fingers(self.obj, vis=False, check_approach=False)
if not grasp_point[0]:
Warning("Could not find contact points")
continue
# Get grasp width in pixel from endpoints
p_1 = np.dot(self.T_obj_camera.matrix, np.append(grasp_point[1][0].point, 1).T).T[:3]
p_2 = np.dot(self.T_obj_camera.matrix, np.append(grasp_point[1][1].point, 1).T).T[:3]
grasp_width_px = self.width_px(shifted_camera_intr, p_1, p_2)
grasp_width = aligned_grasp.width_from_endpoints(grasp_point[1][0].point, grasp_point[1][1].point)
depth_im_no_rot = self.prepare_images(depth_im_table, grasp_2d)
T_grasp_camera = aligned_grasp.gripper_pose(self.gripper).inverse() * self.T_obj_camera.inverse()
if VISUALISE_3D:
z_axis = T_grasp_camera.z_axis
theta = np.rad2deg(np.arccos((z_axis[2]) /
(np.sqrt((z_axis[0] ** 2 + z_axis[1] ** 2 + z_axis[2] ** 2)))))
vis.figure()
T_obj_world = vis.mesh_stable_pose(self.obj.mesh.trimesh,
stable_pose.T_obj_world, style='surface', dim=0.5, plot_table=True)
T_camera_world = T_obj_world * self.T_obj_camera.inverse()
vis.gripper(self.gripper, aligned_grasp, T_obj_world, color=(0.3, 0.3, 0.3),
T_camera_world=T_camera_world)
print(theta)
vis.show()
hand_pose = self.get_hand_pose(grasp_2d, T_grasp_camera, grasp_width, grasp_width_px)
self.add_datapoint(depth_im_no_rot, hand_pose,
collision_free, aligned_grasp, grasp_metrics)
self.cur_image_label += 1
def generate_gqcnn_dataset(dataset_path, database, target_object_keys,
env_rv_params, gripper_name, config):
"""
Generates a GQ-CNN TensorDataset for training models with new grippers, quality metrics, objects, and cameras.
Parameters
----------
dataset_path : str
path to save the dataset to
database : :obj:`Hdf5Database`
Dex-Net database containing the 3D meshes, grasps, and grasp metrics
target_object_keys : :obj:`OrderedDict`
dictionary mapping dataset names to target objects (either 'all' or a list of specific object keys)
env_rv_params : :obj:`OrderedDict`
parameters of the camera and object random variables used in sampling (see meshpy_berkeley.UniformPlanarWorksurfaceImageRandomVariable for more info)
gripper_name : str
name of the gripper to use
config : :obj:`autolab_core.YamlConfig`
other parameters for dataset generation
Notes
-----
Required parameters of config are specified in Other Parameters
Other Parameters
----------------
images_per_stable_pose : int
number of object and camera poses to sample for each stable pose
stable_pose_min_p : float
minimum probability of occurrence for a stable pose to be used in data generation (used to prune bad stable poses
gqcnn/crop_width : int
width, in pixels, of crop region around each grasp center, before resize (changes the size of the region seen by the GQ-CNN)
gqcnn/crop_height : int
height, in pixels, of crop region around each grasp center, before resize (changes the size of the region seen by the GQ-CNN)
gqcnn/final_width : int
width, in pixels, of final transformed grasp image for input to the GQ-CNN (defaults to 32)
gqcnn/final_height : int
height, in pixels, of final transformed grasp image for input to the GQ-CNN (defaults to 32)
table_alignment/max_approach_table_angle : float
max angle between the grasp axis and the table normal when the grasp approach is maximally aligned with the table normal
table_alignment/max_approach_offset : float
max deviation from perpendicular approach direction to use in grasp collision checking
table_alignment/num_approach_offset_samples : int
number of approach samples to use in collision checking
collision_checking/table_offset : float
max allowable interpenetration between the gripper and table to be considered collision free
collision_checking/table_mesh_filename : str
path to a table mesh for collision checking (default data/meshes/table.obj)
collision_checking/approach_dist : float
distance, in meters, between the approach pose and final grasp pose along the grasp axis
collision_checking/delta_approach : float
amount, in meters, to discretize the straight-line path from the gripper approach pose to the final grasp pose
tensors/datapoints_per_file : int
number of datapoints to store in each unique tensor file on disk
tensors/fields : :obj:`dict`
dictionary mapping field names to dictionaries specifying the data type, height, width, and number of channels for each tensor
debug : bool
True (or 1) if the random seed should be set to enforce deterministic behavior, False (0) otherwise
vis/candidate_grasps : bool
True (or 1) if the collision free candidate grasps should be displayed in 3D (for debugging)
vis/rendered_images : bool
True (or 1) if the rendered images for each stable pose should be displayed (for debugging)
vis/grasp_images : bool
True (or 1) if the transformed grasp images should be displayed (for debugging)
"""
# read data gen params
output_dir = dataset_path
gripper = RobotGripper.load(gripper_name)
image_samples_per_stable_pose = config['images_per_stable_pose']
stable_pose_min_p = config['stable_pose_min_p']
# read gqcnn params
gqcnn_params = config['gqcnn']
im_crop_height = gqcnn_params['crop_height']
im_crop_width = gqcnn_params['crop_width']
im_final_height = gqcnn_params['final_height']
im_final_width = gqcnn_params['final_width']
cx_crop = float(im_crop_width) / 2
cy_crop = float(im_crop_height) / 2
# open database
dataset_names = target_object_keys.keys()
datasets = [database.dataset(dn) for dn in dataset_names]
# set target objects
for dataset in datasets:
if target_object_keys[dataset.name] == 'all':
target_object_keys[dataset.name] = dataset.object_keys
# setup grasp params
table_alignment_params = config['table_alignment']
min_grasp_approach_offset = -np.deg2rad(
table_alignment_params['max_approach_offset'])
max_grasp_approach_offset = np.deg2rad(
table_alignment_params['max_approach_offset'])
max_grasp_approach_table_angle = np.deg2rad(
table_alignment_params['max_approach_table_angle'])
num_grasp_approach_samples = table_alignment_params[
'num_approach_offset_samples']
phi_offsets = []
if max_grasp_approach_offset == min_grasp_approach_offset:
phi_inc = 1
elif num_grasp_approach_samples == 1:
phi_inc = max_grasp_approach_offset - min_grasp_approach_offset + 1
else:
phi_inc = (max_grasp_approach_offset - min_grasp_approach_offset) / (
num_grasp_approach_samples - 1)
phi = min_grasp_approach_offset
while phi <= max_grasp_approach_offset:
phi_offsets.append(phi)
phi += phi_inc
# setup collision checking
coll_check_params = config['collision_checking']
approach_dist = coll_check_params['approach_dist']
delta_approach = coll_check_params['delta_approach']
table_offset = coll_check_params['table_offset']
table_mesh_filename = coll_check_params['table_mesh_filename']
if not os.path.isabs(table_mesh_filename):
table_mesh_filename = os.path.join(
os.path.dirname(os.path.realpath(__file__)), '..',
table_mesh_filename)
table_mesh = ObjFile(table_mesh_filename).read()
# set tensor dataset config
tensor_config = config['tensors']
tensor_config['fields']['depth_ims_tf_table']['height'] = im_final_height
tensor_config['fields']['depth_ims_tf_table']['width'] = im_final_width
tensor_config['fields']['obj_masks']['height'] = im_final_height
tensor_config['fields']['obj_masks']['width'] = im_final_width
# add available metrics (assuming same are computed for all objects)
metric_names = []
dataset = datasets[0]
obj_keys = dataset.object_keys
if len(obj_keys) == 0:
raise ValueError('No valid objects in dataset %s' % (dataset.name))
obj = dataset[obj_keys[0]]
grasps = dataset.grasps(obj.key, gripper=gripper.name)
grasp_metrics = dataset.grasp_metrics(obj.key,
grasps,
gripper=gripper.name)
metric_names = grasp_metrics[grasp_metrics.keys()[0]].keys()
for metric_name in metric_names:
tensor_config['fields'][metric_name] = {}
tensor_config['fields'][metric_name]['dtype'] = 'float32'
# init tensor dataset
tensor_dataset = TensorDataset(output_dir, tensor_config)
tensor_datapoint = tensor_dataset.datapoint_template
# setup log file
experiment_log_filename = os.path.join(output_dir,
'dataset_generation.log')
formatter = logging.Formatter('%(asctime)s %(levelname)s: %(message)s')
hdlr = logging.FileHandler(experiment_log_filename)
hdlr.setFormatter(formatter)
logging.getLogger().addHandler(hdlr)
root_logger = logging.getLogger()
# copy config
out_config_filename = os.path.join(output_dir, 'dataset_generation.json')
ordered_dict_config = collections.OrderedDict()
for key in config.keys():
ordered_dict_config[key] = config[key]
with open(out_config_filename, 'w') as outfile:
json.dump(ordered_dict_config, outfile)
# 1. Precompute the set of valid grasps for each stable pose:
# i) Perpendicular to the table
# ii) Collision-free along the approach direction
# load grasps if they already exist
grasp_cache_filename = os.path.join(output_dir, CACHE_FILENAME)
if os.path.exists(grasp_cache_filename):
logging.info('Loading grasp candidates from file')
candidate_grasps_dict = pkl.load(open(grasp_cache_filename, 'rb'))
# otherwise re-compute by reading from the database and enforcing constraints
else:
# create grasps dict
candidate_grasps_dict = {}
# loop through datasets and objects
for dataset in datasets:
logging.info('Reading dataset %s' % (dataset.name))
for obj in dataset:
if obj.key not in target_object_keys[dataset.name]:
continue
# init candidate grasp storage
candidate_grasps_dict[obj.key] = {}
# setup collision checker
collision_checker = GraspCollisionChecker(gripper)
collision_checker.set_graspable_object(obj)
# read in the stable poses of the mesh
stable_poses = dataset.stable_poses(obj.key)
for i, stable_pose in enumerate(stable_poses):
# render images if stable pose is valid
if stable_pose.p > stable_pose_min_p:
candidate_grasps_dict[obj.key][stable_pose.id] = []
# setup table in collision checker
T_obj_stp = stable_pose.T_obj_table.as_frames(
'obj', 'stp')
T_obj_table = obj.mesh.get_T_surface_obj(
T_obj_stp,
delta=table_offset).as_frames('obj', 'table')
T_table_obj = T_obj_table.inverse()
collision_checker.set_table(table_mesh_filename,
T_table_obj)
# read grasp and metrics
grasps = dataset.grasps(obj.key, gripper=gripper.name)
logging.info(
'Aligning %d grasps for object %s in stable %s' %
(len(grasps), obj.key, stable_pose.id))
# align grasps with the table
aligned_grasps = [
grasp.perpendicular_table(stable_pose)
for grasp in grasps
]
# check grasp validity
logging.info(
'Checking collisions for %d grasps for object %s in stable %s'
% (len(grasps), obj.key, stable_pose.id))
for aligned_grasp in aligned_grasps:
# check angle with table plane and skip unaligned grasps
_, grasp_approach_table_angle, _ = aligned_grasp.grasp_angles_from_stp_z(
stable_pose)
perpendicular_table = (
np.abs(grasp_approach_table_angle) <
max_grasp_approach_table_angle)
if not perpendicular_table:
continue
# check whether any valid approach directions are collision free
collision_free = False
for phi_offset in phi_offsets:
rotated_grasp = aligned_grasp.grasp_y_axis_offset(
phi_offset)
collides = collision_checker.collides_along_approach(
rotated_grasp, approach_dist,
delta_approach)
if not collides:
collision_free = True
break
# store if aligned to table
candidate_grasps_dict[obj.key][
stable_pose.id].append(
GraspInfo(aligned_grasp, collision_free))
# visualize if specified
if collision_free and config['vis'][
'candidate_grasps']:
logging.info('Grasp %d' % (aligned_grasp.id))
vis.figure()
vis.gripper_on_object(gripper, aligned_grasp,
obj,
stable_pose.T_obj_world)
vis.show()
# save to file
logging.info('Saving to file')
pkl.dump(candidate_grasps_dict, open(grasp_cache_filename, 'wb'))
# 2. Render a dataset of images and associate the gripper pose with image coordinates for each grasp in the Dex-Net database
# setup variables
obj_category_map = {}
pose_category_map = {}
cur_pose_label = 0
cur_obj_label = 0
cur_image_label = 0
# render images for each stable pose of each object in the dataset
render_modes = [RenderMode.SEGMASK, RenderMode.DEPTH_SCENE]
for dataset in datasets:
logging.info('Generating data for dataset %s' % (dataset.name))
# iterate through all object keys
object_keys = dataset.object_keys
for obj_key in object_keys:
obj = dataset[obj_key]
if obj.key not in target_object_keys[dataset.name]:
continue
# read in the stable poses of the mesh
stable_poses = dataset.stable_poses(obj.key)
for i, stable_pose in enumerate(stable_poses):
# render images if stable pose is valid
if stable_pose.p > stable_pose_min_p:
# log progress
logging.info('Rendering images for object %s in %s' %
(obj.key, stable_pose.id))
# add to category maps
if obj.key not in obj_category_map.keys():
obj_category_map[obj.key] = cur_obj_label
pose_category_map['%s_%s' %
(obj.key,
stable_pose.id)] = cur_pose_label
# read in candidate grasps and metrics
candidate_grasp_info = candidate_grasps_dict[obj.key][
stable_pose.id]
candidate_grasps = [g.grasp for g in candidate_grasp_info]
grasp_metrics = dataset.grasp_metrics(obj.key,
candidate_grasps,
gripper=gripper.name)
# compute object pose relative to the table
T_obj_stp = stable_pose.T_obj_table.as_frames('obj', 'stp')
T_obj_stp = obj.mesh.get_T_surface_obj(T_obj_stp)
# sample images from random variable
T_table_obj = RigidTransform(from_frame='table',
to_frame='obj')
scene_objs = {
'table': SceneObject(table_mesh, T_table_obj)
}
urv = UniformPlanarWorksurfaceImageRandomVariable(
obj.mesh,
render_modes,
'camera',
env_rv_params,
stable_pose=stable_pose,
scene_objs=scene_objs)
render_start = time.time()
render_samples = urv.rvs(
size=image_samples_per_stable_pose)
render_stop = time.time()
logging.info('Rendering images took %.3f sec' %
(render_stop - render_start))
# visualize
if config['vis']['rendered_images']:
d = int(np.ceil(
np.sqrt(image_samples_per_stable_pose)))
# binary
vis2d.figure()
for j, render_sample in enumerate(render_samples):
vis2d.subplot(d, d, j + 1)
vis2d.imshow(render_sample.renders[
RenderMode.SEGMASK].image)
# depth table
vis2d.figure()
for j, render_sample in enumerate(render_samples):
vis2d.subplot(d, d, j + 1)
vis2d.imshow(render_sample.renders[
RenderMode.DEPTH_SCENE].image)
vis2d.show()
# tally total amount of data
num_grasps = len(candidate_grasps)
num_images = image_samples_per_stable_pose
num_save = num_images * num_grasps
logging.info('Saving %d datapoints' % (num_save))
# for each candidate grasp on the object compute the projection
# of the grasp into image space
for render_sample in render_samples:
# read images
binary_im = render_sample.renders[
RenderMode.SEGMASK].image
depth_im_table = render_sample.renders[
RenderMode.DEPTH_SCENE].image
# read camera params
T_stp_camera = render_sample.camera.object_to_camera_pose
shifted_camera_intr = render_sample.camera.camera_intr
# read pixel offsets
cx = depth_im_table.center[1]
cy = depth_im_table.center[0]
# compute intrinsics for virtual camera of the final
# cropped and rescaled images
camera_intr_scale = float(im_final_height) / float(
im_crop_height)
cropped_camera_intr = shifted_camera_intr.crop(
im_crop_height, im_crop_width, cy, cx)
final_camera_intr = cropped_camera_intr.resize(
camera_intr_scale)
# create a thumbnail for each grasp
for grasp_info in candidate_grasp_info:
# read info
grasp = grasp_info.grasp
collision_free = grasp_info.collision_free
# get the gripper pose
T_obj_camera = T_stp_camera * T_obj_stp.as_frames(
'obj', T_stp_camera.from_frame)
grasp_2d = grasp.project_camera(
T_obj_camera, shifted_camera_intr)
# center images on the grasp, rotate to image x axis
dx = cx - grasp_2d.center.x
dy = cy - grasp_2d.center.y
translation = np.array([dy, dx])
binary_im_tf = binary_im.transform(
translation, grasp_2d.angle)
depth_im_tf_table = depth_im_table.transform(
translation, grasp_2d.angle)
# crop to image size
binary_im_tf = binary_im_tf.crop(
im_crop_height, im_crop_width)
depth_im_tf_table = depth_im_tf_table.crop(
im_crop_height, im_crop_width)
# resize to image size
binary_im_tf = binary_im_tf.resize(
(im_final_height, im_final_width),
interp='nearest')
depth_im_tf_table = depth_im_tf_table.resize(
(im_final_height, im_final_width))
# visualize the transformed images
if config['vis']['grasp_images']:
grasp_center = Point(
depth_im_tf_table.center,
frame=final_camera_intr.frame)
tf_grasp_2d = Grasp2D(
grasp_center,
0,
grasp_2d.depth,
width=gripper.max_width,
camera_intr=final_camera_intr)
# plot 2D grasp image
vis2d.figure()
vis2d.subplot(2, 2, 1)
vis2d.imshow(binary_im)
vis2d.grasp(grasp_2d)
vis2d.subplot(2, 2, 2)
vis2d.imshow(depth_im_table)
vis2d.grasp(grasp_2d)
vis2d.subplot(2, 2, 3)
vis2d.imshow(binary_im_tf)
vis2d.grasp(tf_grasp_2d)
vis2d.subplot(2, 2, 4)
vis2d.imshow(depth_im_tf_table)
vis2d.grasp(tf_grasp_2d)
vis2d.title('Coll Free? %d' %
(grasp_info.collision_free))
vis2d.show()
# plot 3D visualization
vis.figure()
T_obj_world = vis.mesh_stable_pose(
obj.mesh,
stable_pose.T_obj_world,
style='surface',
dim=0.5)
vis.gripper(gripper,
grasp,
T_obj_world,
color=(0.3, 0.3, 0.3))
vis.show()
# form hand pose array
hand_pose = np.r_[grasp_2d.center.y,
grasp_2d.center.x,
grasp_2d.depth, grasp_2d.angle,
grasp_2d.center.y -
shifted_camera_intr.cy,
grasp_2d.center.x -
shifted_camera_intr.cx,
grasp_2d.width_px]
# store to data buffers
tensor_datapoint[
'depth_ims_tf_table'] = depth_im_tf_table.raw_data
tensor_datapoint[
'obj_masks'] = binary_im_tf.raw_data
tensor_datapoint['hand_poses'] = hand_pose
tensor_datapoint['collision_free'] = collision_free
tensor_datapoint['obj_labels'] = cur_obj_label
tensor_datapoint['pose_labels'] = cur_pose_label
tensor_datapoint['image_labels'] = cur_image_label
for metric_name, metric_val in grasp_metrics[
grasp.id].iteritems():
coll_free_metric = (
1 * collision_free) * metric_val
tensor_datapoint[
metric_name] = coll_free_metric
tensor_dataset.add(tensor_datapoint)
# update image label
cur_image_label += 1
# update pose label
cur_pose_label += 1
# force clean up
gc.collect()
# update object label
cur_obj_label += 1
# force clean up
gc.collect()
# save last file
tensor_dataset.flush()
# save category mappings
obj_cat_filename = os.path.join(output_dir, 'object_category_map.json')
json.dump(obj_category_map, open(obj_cat_filename, 'w'))
pose_cat_filename = os.path.join(output_dir, 'pose_category_map.json')
json.dump(pose_category_map, open(pose_cat_filename, 'w'))
def benchmark_bin_picking_policy(policy,
# input_dataset_path,
# heap_ids,
# timesteps,
# output_dataset_path,
config,
# excluded_heaps_file):
):
""" Benchmark a bin picking policy.
Parameters
----------
policy : :obj:`Policy`
policy to roll out
input_dataset_path : str
path to the input dataset
heap_ids : list
integer identifiers for the heaps to re-run
timesteps : list
integer timesteps to seed the simulation from
output_dataset_path : str
path to store the results
config : dict
dictionary-like objects containing parameters of the simulator and visualization
"""
# read subconfigs
vis_config = config['vis']
dataset_config = config['dataset']
# read parameters
fully_observed = config['fully_observed']
steps_per_test_case = config['steps_per_test_case']
rollouts_per_garbage_collect = config['rollouts_per_garbage_collect']
debug = config['debug']
im_height = config['state_space']['camera']['im_height']
im_width = config['state_space']['camera']['im_width']
max_obj_per_pile = config['state_space']['object']['max_obj_per_pile']
if debug:
random.seed(SEED)
np.random.seed(SEED)
# read ids
# if len(heap_ids) != len(timesteps):
# raise ValueError('Must provide same number of heap ids and timesteps')
# num_rollouts = len(heap_ids)
num_rollouts = 1
# set dataset params
tensor_config = dataset_config['tensors']
fields_config = tensor_config['fields']
# fields_config['color_ims']['height'] = im_height
# fields_config['color_ims']['width'] = im_width
# fields_config['depth_ims']['height'] = im_height
# fields_config['depth_ims']['width'] = im_width
fields_config['obj_poses']['height'] = POSE_DIM * max_obj_per_pile
fields_config['obj_coms']['height'] = POINT_DIM * max_obj_per_pile
fields_config['obj_ids']['height'] = max_obj_per_pile
fields_config['bin_distances']['height'] = max_obj_per_pile
# matrix has (n choose 2) elements in it
max_distance_matrix_length = int(comb(max_obj_per_pile, 2))
fields_config['distance_matrix']['height'] = max_distance_matrix_length
# sample a process id
proc_id = utils.gen_experiment_id()
# if not os.path.exists(output_dataset_path):
# try:
# os.mkdir(output_dataset_path)
# except:
# logging.warning('Failed to create %s. The dataset path may have been created simultaneously by another process' %(dataset_path))
# proc_id = 'clustering_2'
# output_dataset_path = os.path.join(output_dataset_path, 'dataset_%s' %(proc_id))
# open input dataset
# logging.info('Opening input dataset: %s' % input_dataset_path)
# input_dataset = TensorDataset.open(input_dataset_path)
# open output_dataset
# logging.info('Opening output dataset: %s' % output_dataset_path)
# dataset = TensorDataset(output_dataset_path, tensor_config)
# datapoint = dataset.datapoint_template
# setup logging
# experiment_log_filename = os.path.join(output_dataset_path, 'dataset_generation.log')
# formatter = logging.Formatter('%(asctime)s %(levelname)s: %(message)s')
# hdlr = logging.FileHandler(experiment_log_filename)
# hdlr.setFormatter(formatter)
# logging.getLogger().addHandler(hdlr)
# config.save(os.path.join(output_dataset_path, 'dataset_generation_params.yaml'))
# key mappings
# we add the empty string as a mapping because if you don't evaluate dexnet on the 'before' state of the push
obj_id = 1
obj_ids = {'': 0}
action_ids = {
'ParallelJawGraspAction': 0,
'SuctionGraspAction': 1,
'LinearPushAction': 2
}
# add action ids
reverse_action_ids = utils.reverse_dictionary(action_ids)
# dataset.add_metadata('action_ids', reverse_action_ids)
# perform rollouts
n = 0
rollout_start = time.time()
current_heap_id = None
while n < num_rollouts:
# create env
create_start = time.time()
bin_picking_env = GraspingEnv(config, vis_config)
create_stop = time.time()
logging.info('Creating env took %.3f sec' %(create_stop-create_start))
# perform rollouts
rollouts_remaining = num_rollouts - n
for i in range(min(rollouts_per_garbage_collect, rollouts_remaining)):
# log current rollout
logging.info('\n')
if n % vis_config['log_rate'] == 0:
logging.info('Rollout: %03d' %(n))
try:
# mark rollout status
data_saved = False
num_steps = 0
# read heap id
# heap_id = heap_ids[n]
# timestep = timesteps[n]
# while heap_id == current_heap_id:# or heap_id < 81:#[226, 287, 325, 453, 469, 577, 601, 894, 921]: 26
# n += 1
# heap_id = heap_ids[n]
# timestep = timesteps[n]
push_logger = logging.getLogger('push')
# push_logger.info('~')
# push_logger.info('Heap ID %d' % heap_id)
# current_heap_id = heap_id
# reset env
reset_start = time.time()
# bin_picking_env.reset_from_dataset(input_dataset,
# heap_id,
# timestep)
bin_picking_env.reset()
state = bin_picking_env.state
environment = bin_picking_env.environment
if fully_observed:
observation = None
else:
observation = bin_picking_env.observation
policy.set_environment(environment)
reset_stop = time.time()
# add objects to mapping
for obj_key in state.obj_keys:
if obj_key not in obj_ids.keys():
obj_ids[obj_key] = obj_id
obj_id += 1
push_logger.info(obj_key)
# save id mappings
reverse_obj_ids = utils.reverse_dictionary(obj_ids)
# dataset.add_metadata('obj_ids', reverse_obj_ids)
# store datapoint env params
# datapoint['heap_ids'] = current_heap_id
# datapoint['camera_poses'] = environment.camera.T_camera_world.vec
# datapoint['camera_intrs'] = environment.camera.intrinsics.vec
# datapoint['robot_poses'] = environment.robot.T_robot_world.vec
# render
if vis_config['initial_state']:
vis3d.figure()
bin_picking_env.render_3d_scene()
vis3d.pose(environment.robot.T_robot_world)
vis3d.show(starting_camera_pose=CAMERA_POSE)
# observe
if vis_config['initial_obs']:
vis2d.figure()
vis2d.imshow(observation, auto_subplot=True)
vis2d.show()
# rollout on current satte
done = False
failed = False
# if isinstance(policy, SingulationFullRolloutPolicy):
# policy.reset_num_failed_grasps()
while not done:
if vis_config['step_stats']:
logging.info('Heap ID: %s' % heap_id)
logging.info('Timestep: %s' % bin_picking_env.timestep)
# get action
policy_start = time.time()
if fully_observed:
action = policy.action(state)
else:
action = policy.action(observation)
policy_stop = time.time()
logging.info('Composite Policy took %.3f sec' %(policy_stop-policy_start))
# render scene before
if vis_config['action']:
#gripper = bin_picking_env.gripper(action)
vis3d.figure()
# GRASPINGENV
# bin_picking_env.render_3d_scene(render_camera=False, workspace_objs_wireframe=False)
bin_picking_env.render_3d_scene()
if isinstance(action, GraspAction):
vis3d.gripper(gripper, action.grasp(gripper))
#if isinstance(action, LinearPushAction):
else:
# # T_start_world = action.T_begin_world * gripper.T_mesh_grasp
# # T_end_world = action.T_end_world * gripper.T_mesh_grasp
# #start_point = action.T_begin_world.translation
# start_point = action['start']
# #end_point = action.T_end_world.translation
# end_point = action['end']
# vec = (end_point - start_point) / np.linalg.norm(end_point-start_point) if np.linalg.norm(end_point-start_point) > 0 else end_point-start_point
# #h1 = np.array([[0.7071,-0.7071,0],[0.7071,0.7071,0],[0,0,1]]).dot(vec)
# #h2 = np.array([[0.7071,0.7071,0],[-0.7071,0.7071,0],[0,0,1]]).dot(vec)
# arrow_len = np.linalg.norm(start_point - end_point)
# h1 = (end_point - start_point + np.array([0,0,arrow_len])) / (arrow_len*math.sqrt(2))
# h2 = (end_point - start_point - np.array([0,0,arrow_len])) / (arrow_len*math.sqrt(2))
# shaft_points = [start_point, end_point]
# head_points = [end_point - 0.03*h2, end_point, end_point - 0.03*h1]
# #vis3d.plot3d(shaft_points, color=[0,0,1])
# #vis3d.plot3d(head_points, color=[0,0,1])
# Displaying all potential topple points
for vertex, prob in zip(action['vertices'], action['probabilities']):
color = np.array([min(1, 2*(1-prob)), min(2*prob, 1), 0])
vis3d.points(Point(vertex, 'world'), scale=.0005, color=color)
for vertex in action['bottom_points']:
color = np.array([0,0,1])
vis3d.points(Point(vertex, 'world'), scale=.0005, color=color)
vis3d.points(Point(action['com'], 'world'), scale=.005, color=np.array([0,0,1]))
vis3d.points(Point(np.array([0,0,0]), 'world'), scale=.005, color=np.array([0,1,0]))
#set_of_lines = action['set_of_lines']
#for i, line in enumerate(set_of_lines):
# color = str(bin(i+1))[2:].zfill(3)
# color = np.array([color[2], color[1], color[0]])
# vis3d.plot3d(line, color=color)
vis3d.show(starting_camera_pose=CAMERA_POSE)
# Show
vis3d.figure()
bin_picking_env.render_3d_scene()
final_pose_ind = action['final_pose_ind'] / np.amax(action['final_pose_ind'])
for vertex, final_pose_ind in zip(action['vertices'], final_pose_ind):
color = np.array([0, min(1, 2*(1-prob)), min(2*prob, 1)])
vis3d.points(Point(vertex, 'world'), scale=.0005, color=color)
vis3d.show(starting_camera_pose=CAMERA_POSE)
color=np.array([0,0,1])
original_pose = state.obj.T_obj_world
pose_num = 0
for pose, edge_point1, edge_point2 in zip(action['final_poses'], action['bottom_points'], np.roll(action['bottom_points'],-1,axis=0)):
print 'Pose:', pose_num
pose_num += 1
pose = pose.T_obj_table
vis3d.figure()
state.obj.T_obj_world = original_pose
bin_picking_env.render_3d_scene()
vis3d.points(Point(edge_point1, 'world'), scale=.0005, color=color)
vis3d.points(Point(edge_point2, 'world'), scale=.0005, color=color)
vis3d.show(starting_camera_pose=CAMERA_POSE)
vis3d.figure()
state.obj.T_obj_world = pose
bin_picking_env.render_3d_scene()
vis3d.points(Point(edge_point1, 'world'), scale=.0005, color=color)
vis3d.points(Point(edge_point2, 'world'), scale=.0005, color=color)
vis3d.show(starting_camera_pose=CAMERA_POSE)
#vis3d.save('/home/mjd3/Pictures/weird_pics/%d_%d_before.png' % (heap_id, bin_picking_env.timestep), starting_camera_pose=CAMERA_POSE)
# store datapoint pre-step data
j = 0
obj_poses = np.zeros(fields_config['obj_poses']['height'])
obj_coms = np.zeros(fields_config['obj_coms']['height'])
obj_ids_vec = np.iinfo(np.uint32).max * np.ones(fields_config['obj_ids']['height'])
for obj_state in state.obj_states:
obj_poses[j*POSE_DIM:(j+1)*POSE_DIM] = obj_state.T_obj_world.vec
obj_coms[j*POINT_DIM:(j+1)*POINT_DIM] = obj_state.center_of_mass
obj_ids_vec[j] = obj_ids[obj_state.key]
j += 1
action_poses = np.zeros(fields_config['action_poses']['height'])
#if isinstance(action, GraspAction):
# action_poses[:7] = action.T_grasp_world.vec
#else:
# action_poses[:7] = action.T_begin_world.vec
# action_poses[7:] = action.T_end_world.vec
# if isinstance(policy, SingulationMetricsCompositePolicy):
# actual_distance_matrix_length = int(comb(len(state.objs), 2))
# bin_distances = np.append(action.metadata['bin_distances'],
# np.zeros(max_obj_per_pile-len(state.objs))
# )
# distance_matrix = np.append(action.metadata['distance_matrix'],
# np.zeros(max_distance_matrix_length - actual_distance_matrix_length)
# )
# datapoint['bin_distances'] = bin_distances
# datapoint['distance_matrix'] = distance_matrix
# datapoint['T_begin_world'] = action.T_begin_world.matrix
# datapoint['T_end_world'] = action.T_end_world.matrix
# datapoint['parallel_jaw_best_q_value'] = action.metadata['parallel_jaw_best_q_value']
# # datapoint['parallel_jaw_mean_q_value'] = action.metadata['parallel_jaw_mean_q_value']
# # datapoint['parallel_jaw_num_grasps'] = action.metadata['parallel_jaw_num_grasps']
# datapoint['suction_best_q_value'] = action.metadata['suction_best_q_value']
# # datapoint['suction_mean_q_value'] = action.metadata['suction_mean_q_value']
# # datapoint['suction_num_grasps'] = action.metadata['suction_num_grasps']
# # logging.info('Suction Q: %f, PJ Q: %f' % (action.metadata['suction_q_value'], action.metadata['parallel_jaw_q_value']))
# # datapoint['obj_index'] = action.metadata['obj_index']
# # datapoint['parallel_jaw_best_q_value_single'] = action.metadata['parallel_jaw_best_q_value_single']
# # datapoint['suction_best_q_value_single'] = action.metadata['suction_best_q_value_single']
# datapoint['singulated_obj_index'] = action.metadata['singulated_obj_index']
# datapoint['parallel_jaw_grasped_obj_index'] = obj_ids[action.metadata['parallel_jaw_grasped_obj_key']]
# datapoint['suction_grasped_obj_index'] = obj_ids[action.metadata['suction_grasped_obj_key']]
# else:
# datapoint['bin_distances'] = np.zeros(max_obj_per_pile)
# datapoint['distance_matrix'] = np.zeros(max_distance_matrix_length)
# datapoint['T_begin_world'] = np.zeros((4,4))
# datapoint['T_end_world'] = np.zeros((4,4))
# datapoint['parallel_jaw_best_q_value'] = -1
# datapoint['suction_best_q_value'] = -1
# datapoint['singulated_obj_index'] = -1
# datapoint['parallel_jaw_grasped_obj_index'] = -1
# datapoint['suction_grasped_obj_index'] = -1
# policy_id = 0
# if 'policy_id' in action.metadata.keys():
# policy_id = action.metadata['policy_id']
# greedy_q_value = 0
# if 'greedy_q_value' in action.metadata.keys():
# greedy_q_value = action.metadata['greedy_q_value']
# datapoint['timesteps'] = bin_picking_env.timestep
# datapoint['obj_poses'] = obj_poses
# datapoint['obj_coms'] = obj_coms
# datapoint['obj_ids'] = obj_ids_vec
# # if bin_picking_env.render_mode == RenderMode.RGBD:
# # color_data = observation.color.raw_data
# # depth_data = observation.depth.raw_data
# # elif bin_picking_env.render_mode == RenderMode.DEPTH:
# # color_data = np.zeros(observation.shape).astype(np.uint8)
# # depth_data = observation.raw_data
# # elif bin_picking_env.render_mode == RenderMode.COLOR:
# # color_data = observation.raw_data
# # depth_data = np.zeros(observation.shape)
# # datapoint['color_ims'] = color_data
# # datapoint['depth_ims'] = depth_data
# datapoint['action_ids'] = action_ids[type(action).__name__]
# datapoint['action_poses'] = action_poses
# datapoint['policy_ids'] = policy_id
# datapoint['greedy_q_values'] = greedy_q_value
# datapoint['pred_q_values'] = action.q_value
# step the policy
#observation, reward, done, info = bin_picking_env.step(action)
#state = bin_picking_env.state
state.objs[0].T_obj_world = action['final_state']
# if isinstance(policy, SingulationFullRolloutPolicy):
# policy.grasp_succeeds(info['grasp_succeeds'])
# debugging info
if vis_config['step_stats']:
logging.info('Action type: %s' %(type(action).__name__))
logging.info('Action Q-value: %.3f' %(action.q_value))
logging.info('Reward: %d' %(reward))
logging.info('Policy took %.3f sec' %(policy_stop-policy_start))
logging.info('Num objects remaining: %d' %(bin_picking_env.num_objects))
if info['cleared_pile']:
logging.info('Cleared pile!')
# # store datapoint post-step data
# datapoint['rewards'] = reward
# datapoint['grasp_metrics'] = info['grasp_metric']
# datapoint['collisions'] = 1 * info['collides']
# datapoint['collisions_with_env'] = 1 * info['collides_with_static_obstacles']
# datapoint['grasped_obj_ids'] = obj_ids[info['grasped_obj_key']]
# datapoint['cleared_pile'] = 1 * info['cleared_pile']
# # store datapoint
# # dataset.add(datapoint)
# data_saved = True
# render observation
if vis_config['obs']:
vis2d.figure()
vis2d.imshow(observation, auto_subplot=True)
vis2d.show()
# render scene after
if vis_config['state']:
vis3d.figure()
bin_picking_env.render_3d_scene(render_camera=False)
vis3d.show(starting_camera_pose=CAMERA_POSE)
# vis3d.save('/home/mjd3/Pictures/weird_pics/%d_%d_after.png' % (heap_id, bin_picking_env.timestep), starting_camera_pose=CAMERA_POSE)
state.objs[0].T_obj_world = action['tmpR']
vis3d.figure()
bin_picking_env.render_3d_scene(render_camera=False)
vis3d.show(starting_camera_pose=CAMERA_POSE)
state.objs[0].T_obj_world = action['final_state']
# increment the number of steps
num_steps += 1
if num_steps >= steps_per_test_case:
done = True
except NoActionFoundException as e:
logging.warning('The policy failed to plan an action!')
done = True
except Exception as e:
# log an error
logging.warning('Rollout failed!')
logging.warning('%s' %(str(e)))
logging.warning(traceback.print_exc())
# if debug:
# raise
# reset env
del bin_picking_env
gc.collect()
bin_picking_env = BinPickingEnv(config, vis_config)
# terminate current rollout
failed = True
done = True
# update test case id
n += 1
# dataset.flush()
# logging.info("\n\nflushing")
# logging.info("exiting")
# sys.exit()
# garbage collect
del bin_picking_env
gc.collect()
# return the dataset
# dataset.flush()
# log time
rollout_stop = time.time()
logging.info('Rollouts took %.3f sec' %(rollout_stop-rollout_start))
return dataset
def save_samples(self, sample, grasps, T_obj_stp, collision_checker,
grasp_metrics, stable_pose):
self.tensor_datapoint['image_labels'] = self.cur_image_label
T_stp_camera = sample.camera.object_to_camera_pose
self.T_obj_camera = T_stp_camera * T_obj_stp.as_frames(
'obj', T_stp_camera.from_frame)
if UNALIGNED or STP_ALIGNED:
aligned_grasps = grasps
else:
aligned_grasps = self.align_grasps_with_camera(grasps)
depth_im_table = sample.renders[RenderMode.DEPTH_SCENE].image
self.tensor_datapoint['camera_poses'] = self.get_camera_pose(sample)
final_camera_intr = sample.camera.camera_intr.crop(
self.crop_height, self.crop_width, depth_im_table.center[0],
depth_im_table.center[1])
self.tensor_datapoint['camera_intrs'] = np.r_[final_camera_intr.fx,
final_camera_intr.fy,
final_camera_intr.cx,
final_camera_intr.cy]
if not IM_TENSOR:
self.save_image(depth_im_table)
else:
depth_im_tf = self._crop(depth_im_table, size=None)
for cnt, aligned_grasp in enumerate(aligned_grasps):
if UNALIGNED:
aligned_grasp = aligned_grasp.grasp_y_axis_offset(
np.random.uniform(0, 2 * np.pi))
T_stp_grasp = RigidTransform(
stable_pose.r, from_frame='obj',
to_frame='stp') * aligned_grasp.T_grasp_obj
if T_stp_grasp.x_axis[-1] > 0:
# Grasp from below table, rotate by 180 degrees
aligned_grasp = aligned_grasp.grasp_y_axis_offset(np.pi)
T_grasp_camera = aligned_grasp.gripper_pose(self.gripper).inverse() * \
self.T_obj_camera.inverse()
# Calculate psi, the angle between the grasp approach axis and the camera principal ray
z_axis = T_grasp_camera.z_axis
psi = np.arccos(
z_axis[2] /
np.sqrt(z_axis[0]**2 + z_axis[1]**2 + z_axis[2]**2))
# if np.rad2deg(psi) > 20:
# continue
# Get gamma, the angle between the grasp approach axis and the table normal
_, gamma, _ = aligned_grasp.grasp_angles_from_stp_z(stable_pose)
if LIMIT_BETA:
perpendicular_table = (np.abs(gamma) < np.deg2rad(5))
if not perpendicular_table:
continue
if VISUALISE_3D:
vis.figure()
T_obj_world = vis.mesh_stable_pose(self.obj.mesh.trimesh,
stable_pose.T_obj_world,
style='surface',
dim=0.5,
plot_table=True)
T_camera_world = T_obj_world * self.T_obj_camera.inverse()
vis.gripper(self.gripper,
aligned_grasp,
T_obj_world,
color=(0.3, 0.3, 0.3),
T_camera_world=T_camera_world)
# vis.gripper(self.gripper, grasps[cnt], T_obj_world, color=(0.3, 0.3, 0.3))
# print(psi)
vis.show()
collision_free = self.is_grasp_collision_free(
aligned_grasp, collision_checker)
# Project grasp coordinates in image
grasp_2d = aligned_grasp.project_camera(self.T_obj_camera,
final_camera_intr)
if IM_TENSOR:
im, new_grasp = self.random_im_crop(depth_im_tf, (32, 32),
grasp_2d.center.x,
grasp_2d.center.y)
self.tensor_datapoint['depth_ims_tf_table'] = im
hand_pose = self.get_hand_pose(grasp_2d,
T_grasp_camera.quaternion,
psi,
gamma,
new_grasp=new_grasp)
else:
hand_pose = self.get_hand_pose(grasp_2d,
T_grasp_camera.quaternion, psi,
gamma)
self.add_datapoint(hand_pose, collision_free, aligned_grasp,
grasp_metrics)
self.cur_image_label += 1
def save_samples(self, sample, grasps, T_obj_stp, collision_checker,
grasp_metrics, stable_pose):
self.tensor_datapoint['image_labels'] = self.cur_image_label
T_stp_camera = sample.camera.object_to_camera_pose
self.T_obj_camera = T_stp_camera * T_obj_stp.as_frames(
'obj', T_stp_camera.from_frame)
depth_im_table = sample.renders[RenderMode.DEPTH_SCENE].image
camera_pose = self.get_camera_pose(sample)
shifted_camera_intr = sample.camera.camera_intr
aligned_grasps = self.align_grasps_with_camera(grasps)
cx = depth_im_table.center[1]
cy = depth_im_table.center[0]
self.tensor_datapoint['camera_intrs'] = np.r_[shifted_camera_intr.fx,
shifted_camera_intr.fy,
shifted_camera_intr.cx,
shifted_camera_intr.cy]
for aligned_grasp in aligned_grasps:
if not self.is_grasp_aligned(aligned_grasp):
continue
collision_free = self.is_grasp_collision_free(
aligned_grasp, collision_checker)
grasp_2d = aligned_grasp.project_camera(self.T_obj_camera,
shifted_camera_intr)
depth_im_tf = self.prepare_images(depth_im_table, grasp_2d, cx, cy)
T_grasp_camera = aligned_grasp.gripper_pose(
self.gripper).inverse() * self.T_obj_camera.inverse()
if VISUALISE_3D:
z_axis = T_grasp_camera.z_axis
theta = np.rad2deg(
np.arccos((z_axis[2]) / (np.sqrt(
(z_axis[0]**2 + z_axis[1]**2 + z_axis[2]**2)))))
vis.figure()
T_obj_world = vis.mesh_stable_pose(self.obj.mesh.trimesh,
stable_pose.T_obj_world,
style='surface',
dim=0.5,
plot_table=True)
T_camera_world = T_obj_world * self.T_obj_camera.inverse()
vis.gripper(self.gripper,
aligned_grasp,
T_obj_world,
color=(0.3, 0.3, 0.3),
T_camera_world=T_camera_world)
vis.show()
theta = grasp_2d.angle
R = np.array([[np.cos(theta), -np.sin(theta), 0],
[np.sin(theta), np.cos(theta), 0], [0, 0, 1]])
rot = RigidTransform(rotation=R,
to_frame='camera',
from_frame='camera')
new_quaternion = T_grasp_camera * rot
hand_pose = self.get_hand_pose(grasp_2d, new_quaternion.quaternion)
self.add_datapoint(depth_im_tf, hand_pose, collision_free,
camera_pose, aligned_grasp, grasp_metrics)
self.cur_image_label += 1
def save_samples(self, sample, grasps, T_obj_stp, collision_checker, grasp_metrics, stable_pose):
self.tensor_datapoint['image_labels'] = self.cur_image_label
T_stp_camera = sample.camera.object_to_camera_pose
self.T_obj_camera = T_stp_camera * T_obj_stp.as_frames('obj', T_stp_camera.from_frame)
aligned_grasps = self.align_grasps_with_camera(grasps)
depth_im_table = sample.renders[RenderMode.DEPTH_SCENE].image
cx = depth_im_table.center[1]
cy = depth_im_table.center[0]
camera_pose = self.get_camera_pose(sample)
self.tensor_datapoint['camera_poses'] = camera_pose
shifted_camera_intr = sample.camera.camera_intr
self.tensor_datapoint['camera_intrs'] = np.r_[shifted_camera_intr.fx,
shifted_camera_intr.fy,
shifted_camera_intr.cx,
shifted_camera_intr.cy]
self.save_orig_image(depth_im_table, camera_pose, stable_pose, shifted_camera_intr)
grid_space = 200
grids = 3
centre = np.empty((grids, grids, 2))
saved = np.ones((grids, grids)) * -1
depth_images = np.empty((grids, grids, 32, 32, 1))
# Crop images in a grid around the centre
for x_grid in range(grids):
x_pos = grid_space / 4 - x_grid * grid_space / (grids + 1)
for y_grid in range(grids):
y_pos = grid_space / 4 - y_grid * grid_space / (grids + 1)
depth_im = self.prepare_images(depth_im_table, x_pos, y_pos)
self._show_image(depth_im)
depth_images[x_grid, y_grid, :, :, :] = depth_im
centre[x_grid, y_grid, 0] = x_pos
centre[x_grid, y_grid, 1] = y_pos
for cnt, aligned_grasp in enumerate(aligned_grasps):
collision_free = self.is_grasp_collision_free(aligned_grasp, collision_checker)
# Project grasp coordinates in image
grasp_2d = aligned_grasp.project_camera(self.T_obj_camera, shifted_camera_intr)
grasp_point = aligned_grasp.close_fingers(self.obj, vis=False, check_approach=False)
if not grasp_point[0]:
Warning("Could not find contact points")
continue
# Get grasp width in pixel from endpoints
p_1 = np.dot(self.T_obj_camera.matrix, np.append(grasp_point[1][0].point, 1).T).T[:3]
p_2 = np.dot(self.T_obj_camera.matrix, np.append(grasp_point[1][1].point, 1).T).T[:3]
grasp_width_px = self.width_px(shifted_camera_intr, p_1, p_2)
grasp_width = aligned_grasp.width_from_endpoints(grasp_point[1][0].point, grasp_point[1][1].point)
pos_x = cx - grasp_2d.center.x
pos_y = cy - grasp_2d.center.y
distance = np.abs((centre[:, :, 0] - pos_x)**2 + (centre[:, :, 1] - pos_y)**2)
idx = np.where(distance == np.amin(distance))
idx_x = idx[0][0]
idx_y = idx[1][0]
grasp_center_x = (300 - grasp_2d.center.x - centre[idx_x, idx_y, 0])//3
grasp_center_y = (300 - grasp_2d.center.y - centre[idx_x, idx_y, 1])//3
if saved[idx_x, idx_y] == -1:
np.save(self.image_dir + "/depth_im_{:07d}.npy".format(self.cur_image_file),
depth_images[idx_x, idx_y, :, :, :])
saved[idx_x, idx_y] = self.cur_image_file
self.tensor_datapoint['image_files'] = self.cur_image_file
self.cur_image_file += 1
else:
self.tensor_datapoint['image_files'] = saved[idx_x, idx_y]
T_grasp_camera = aligned_grasp.gripper_pose(self.gripper).inverse() * self.T_obj_camera.inverse()
if VISUALISE_3D:
z_axis = T_grasp_camera.z_axis
theta = np.rad2deg(np.arccos((z_axis[2]) /
(np.sqrt((z_axis[0] ** 2 + z_axis[1] ** 2 + z_axis[2] ** 2)))))
vis.figure()
T_obj_world = vis.mesh_stable_pose(self.obj.mesh.trimesh,
stable_pose.T_obj_world, style='surface', dim=0.5, plot_table=True)
T_camera_world = T_obj_world * self.T_obj_camera.inverse()
vis.gripper(self.gripper, aligned_grasp, T_obj_world, color=(0.3, 0.3, 0.3),
T_camera_world=T_camera_world)
print(theta)
vis.show()
hand_pose = self.get_hand_pose(grasp_center_x, grasp_center_y, grasp_2d, T_grasp_camera,
grasp_width, grasp_width_px)
self.add_datapoint(hand_pose,
collision_free, aligned_grasp, grasp_metrics)
self.cur_image_label += 1
if not collides:
collision_free = True
break
# Load grasp metrics
grasp_metrics = dataset.grasp_metrics(obj.key, aligned_grasps, gripper=gripper.name)
# Project grasp coordinates in image
grasp_2d = aligned_grasp.project_camera(T_obj_camera, shifted_camera_intr)
if VISUALISE_3D:
vis.figure()
T_obj_world = vis.mesh_stable_pose(obj.mesh.trimesh,
stable_pose.T_obj_world, style='surface', dim=0.5)
T_camera_world = T_obj_world * T_obj_camera.inverse()
vis.gripper(gripper, aligned_grasp, T_obj_world, color=(0.3, 0.3, 0.3),
T_camera_world=T_camera_world)
vis.show()
# Get translation image distances to grasp
dx = cx - grasp_2d.center.x
dy = cy - grasp_2d.center.y
translation = np.array([dy, dx])
# Transform, crop and resize image
binary_im_tf = binary_im.transform(translation, grasp_2d.angle)
depth_im_tf_table = depth_im_table.transform(translation, grasp_2d.angle)
binary_im_tf = binary_im_tf.crop(96, 96)
depth_im_tf_table = depth_im_tf_table.crop(96, 96)
depth_image = np.asarray(depth_im_tf_table.data)
def save_samples(self, sample, grasps, collisions, T_obj_stp,
grasp_metrics, stable_pose):
self.tensor_datapoint['image_labels'] = self.cur_image_label
T_stp_camera = sample.camera.object_to_camera_pose
self.T_obj_camera = T_stp_camera * T_obj_stp.as_frames(
'obj', T_stp_camera.from_frame)
depth_im_table = sample.renders[RenderMode.DEPTH_SCENE].image
camera_pose = self.get_camera_pose(sample)
final_camera_intr = sample.camera.camera_intr.crop(
self.crop_height, self.crop_width, depth_im_table.center[0],
depth_im_table.center[1])
self.tensor_datapoint['camera_intrs'] = np.r_[final_camera_intr.fx,
final_camera_intr.fy,
final_camera_intr.cx,
final_camera_intr.cy]
self.save_image(depth_im_table)
for aligned_grasp, collision_free in zip(grasps, collisions):
# Project grasp coordinates in image
grasp_point = aligned_grasp.close_fingers(self.obj,
vis=False,
check_approach=False)
if not grasp_point[0]:
Warning("Could not find contact points")
continue
# Get grasp width in pixel from endpoints
p_1 = np.dot(self.T_obj_camera.matrix,
np.append(grasp_point[1][0].point, 1).T).T[:3]
p_2 = np.dot(self.T_obj_camera.matrix,
np.append(grasp_point[1][1].point, 1).T).T[:3]
grasp_width_px = self.width_px(final_camera_intr, p_1, p_2)
grasp_width = aligned_grasp.width_from_endpoints(
grasp_point[1][0].point, grasp_point[1][1].point)
if VISUALISE_3D:
c = np.array((1, 0, 0))
vis.figure()
T_obj_world = vis.mesh_stable_pose(self.obj.mesh.trimesh,
stable_pose.T_obj_world,
style='surface',
dim=0.5,
plot_table=True)
points = np.array(
(np.append(grasp_point[1][0].point,
1), np.append(grasp_point[1][1].point, 1)))
vis.gripper(self.gripper,
aligned_grasp,
T_obj_world,
color=(0.3, 0.3, 0.3),
point=points,
point_color=c)
vis.show()
grasp_2d = aligned_grasp.project_camera(self.T_obj_camera,
final_camera_intr)
grasp_center_x = self.crop_width // 2 - grasp_2d.center.x
grasp_center_y = self.crop_height // 2 - grasp_2d.center.y
hand_pose = self.get_hand_pose(grasp_center_x, grasp_center_y,
grasp_2d, final_camera_intr,
grasp_width, grasp_width_px)
self.add_datapoint(hand_pose, collision_free, camera_pose,
aligned_grasp, grasp_metrics)
self.cur_image_label += 1
def save_samples(self, sample, grasps, collisions, T_obj_stp,
grasp_metrics, stable_pose):
self.tensor_datapoint['image_labels'] = self.cur_image_label
T_stp_camera = sample.camera.object_to_camera_pose
self.T_obj_camera = T_stp_camera * T_obj_stp.as_frames(
'obj', T_stp_camera.from_frame)
depth_im_table = sample.renders[RenderMode.DEPTH_SCENE].image
camera_pose = self.get_camera_pose(sample)
shifted_camera_intr = sample.camera.camera_intr
cx = depth_im_table.center[1]
cy = depth_im_table.center[0]
self.tensor_datapoint['camera_intrs'] = np.r_[shifted_camera_intr.fx,
shifted_camera_intr.fy,
shifted_camera_intr.cx,
shifted_camera_intr.cy]
for aligned_grasp, collision_free in zip(grasps, collisions):
# Project grasp coordinates in image
grasp_point = aligned_grasp.close_fingers(self.obj,
vis=False,
check_approach=False)
if not grasp_point[0]:
raise Warning("Could not find contact points")
continue
grasp_width = aligned_grasp.width_from_endpoints(
grasp_point[1][0].point, grasp_point[1][1].point)
if VISUALISE_3D:
c = np.array((1, 0, 0))
vis.figure()
T_obj_world = vis.mesh_stable_pose(self.obj.mesh.trimesh,
stable_pose.T_obj_world,
style='surface',
dim=0.5,
plot_table=True)
points = np.array(
(np.append(grasp_point[1][0].point,
1), np.append(grasp_point[1][1].point, 1)))
vis.gripper(self.gripper,
aligned_grasp,
T_obj_world,
color=(0.3, 0.3, 0.3),
point=points,
point_color=c)
vis.show()
grasp_2d = aligned_grasp.project_camera(self.T_obj_camera,
shifted_camera_intr)
depth_im_tf = self.prepare_images(depth_im_table, grasp_2d, cx, cy)
hand_pose = self.get_hand_pose(grasp_2d, shifted_camera_intr,
grasp_width)
self.add_datapoint(depth_im_tf, hand_pose, collision_free,
camera_pose, aligned_grasp, grasp_metrics)
self.cur_image_label += 1