def main(scene_idx=0):
#scene_idx = 0
mapper_scene2z = get_mapper()
mapper_scene2points = get_mapper_scene2points()
Train_Scenes, Test_Scenes = get_train_test_scenes()
scene_name = Test_Scenes[scene_idx]
num_startPoints = len(mapper_scene2points[scene_name])
num_steps = 35
## create test folder
test_folder = '/home/reza/Datasets/GibsonEnv/my_code/visual_servoing/test_IBVS'
'''
approach_folder = '{}/SIFT_interMatrix_gtDepth_Vz_OmegaY'.format(test_folder)
create_folder(approach_folder)
scene_folder = '{}/{}'.format(approach_folder, scene_name)
create_folder(scene_folder)
'''
f = open('{}/{}_{}.txt'.format(test_folder, perception_rep, depth_method), 'a')
f.write('scene_name = {}\n'.format(scene_name))
list_count_correct = []
list_count_runs = []
list_count_steps = []
## rrt functions
## first figure out how to sample points from rrt graph
rrt_directory = '/home/reza/Datasets/GibsonEnv/gibson/assets/dataset/{}_for_rrt'.format(scene_name)
path_finder = rrt.PathFinder(rrt_directory)
path_finder.load()
num_nodes = len(path_finder.nodes_x)
free = cv2.imread('/home/reza/Datasets/GibsonEnv/gibson/assets/dataset/{}_for_rrt/free.png'.format(scene_name), 0)
## GibsonEnv setup
## For Gibson Env
import gym, logging
from mpi4py import MPI
from gibson.envs.husky_env import HuskyNavigateEnv
from baselines import logger
import skimage.io
from transforms3d.euler import euler2quat
config_file = os.path.join('/home/reza/Datasets/GibsonEnv/my_code/CVPR_workshop', 'env_yamls', '{}_navigate.yaml'.format(scene_name))
env = HuskyNavigateEnv(config=config_file, gpu_count = 1)
obs = env.reset() ## this line is important otherwise there will be an error like 'AttributeError: 'HuskyNavigateEnv' object has no attribute 'potential''
def get_obs(current_pose):
pos, orn = func_pose2posAndorn(current_pose, mapper_scene2z[scene_name])
env.robot.reset_new_pose(pos, orn)
obs, _, _, _ = env.step(4)
obs_rgb = obs['rgb_filled']
obs_depth = obs['depth']
return obs_rgb.copy(), obs_depth.copy()
base_folder = '/home/reza/Datasets/GibsonEnv/my_code/visual_servoing/sample_image_pairs_{}'.format('test')
## go through each point folder
for point_idx in range(0, num_startPoints):
#for point_idx in range(0, 1):
print('point_idx = {}'.format(point_idx))
#point_folder = '{}/point_{}'.format(scene_folder, point_idx)
#create_folder(point_folder)
## read in start img and start pose
point_image_folder = '{}/{}/point_{}'.format(base_folder, scene_name, point_idx)
point_pose_npy_file = np.load('{}/{}/point_{}_poses.npy'.format(base_folder, scene_name, point_idx))
start_img = cv2.imread('{}/{}.png'.format(point_image_folder, point_pose_npy_file[0]['img_name']))[:, :, ::-1]
start_pose = point_pose_npy_file[0]['pose']
## index 0 is the left image, so right_img_idx starts from index 1
count_correct = 0
list_correct_img_names = []
list_whole_stat = []
list_steps = []
for right_img_idx in range(1, len(point_pose_npy_file)):
#for right_img_idx in range(1, 11):
flag_correct = False
print('right_img_idx = {}'.format(right_img_idx))
count_steps = 0
current_pose = start_pose
right_img_name = point_pose_npy_file[right_img_idx]['img_name']
goal_pose = point_pose_npy_file[right_img_idx]['pose']
goal_img, goal_depth = get_obs(goal_pose)
list_result_poses = [current_pose]
num_matches = 0
flag_broken = False
while count_steps < num_steps:
current_img, current_depth = get_obs(current_pose)
noise = np.random.normal(0.0, 0.5, (256, 256))
current_depth[:, :, 0] += noise
try:
kp1, kp2 = detect_correspondences(current_img, goal_img)
if count_steps == 0:
start_depth = current_depth.copy()
except:
print('run into error')
break
num_matches = kp1.shape[1]
vx, vz, omegay, flag_stop = compute_velocity_through_correspondences_and_depth(kp1, kp2, current_depth)
previous_pose = current_pose
current_pose, _, _, flag_stop_goToPose = goToPose_one_step(current_pose, vx, vz, omegay)
## check if there is collision during the action
left_pixel = path_finder.point_to_pixel((previous_pose[0], previous_pose[1]))
right_pixel = path_finder.point_to_pixel((current_pose[0], current_pose[1]))
# rrt.line_check returns True when there is no obstacle
if not rrt.line_check(left_pixel, right_pixel, free):
flag_broken = True
print('run into an obstacle ...')
break
## check if we should stop or not
if flag_stop or flag_stop_goToPose:
print('flag_stop = {}, flag_stop_goToPose = {}'.format(flag_stop, flag_stop_goToPose))
print('break')
break
count_steps += 1
list_result_poses.append(current_pose)
## sample current_img again to save in list_obs
current_img, current_depth = get_obs(current_pose)
#assert 1==2
## decide if this run is successful or not
flag_correct, dist, theta_change = similar_location_under_certainThreshold(goal_pose, list_result_poses[count_steps])
print('dist = {}, theta = {}'.format(dist, theta_change))
#print('start_pose = {}, final_pose = {}, goal_pose = {}'.format(start_pose, list_result_poses[-1], goal_pose))
if flag_correct:
count_correct += 1
list_correct_img_names.append(right_img_name[10:])
if flag_correct:
str_succ = 'Success'
print('str_succ = {}'.format(str_succ))
else:
str_succ = 'Failure'
print('str_succ = {}'.format(str_succ))
list_steps.append(len(list_result_poses))
## ===================================================================================================================
## plot the pose graph
'''
img_name = 'goTo_{}.jpg'.format(right_img_name[10:])
print('img_name = {}'.format(img_name))
## plot the poses
free2 = cv2.imread('/home/reza/Datasets/GibsonEnv/gibson/assets/dataset/{}_for_rrt/free.png'.format(scene_name), 1)
rows, cols, _ = free2.shape
plt.imshow(free2)
for m in range(len(list_result_poses)):
pose = list_result_poses[m]
x, y = path_finder.point_to_pixel((pose[0], pose[1]))
theta = pose[2]
plt.arrow(x, y, cos(theta), sin(theta), color='y', \
overhang=1, head_width=0.1, head_length=0.15, width=0.001)
## draw goal pose
x, y = path_finder.point_to_pixel((goal_pose[0], goal_pose[1]))
theta = goal_pose[2]
plt.arrow(x, y, cos(theta), sin(theta), color='r', \
overhang=1, head_width=0.1, head_length=0.15, width=0.001)
plt.axis([0, cols, 0, rows])
plt.xticks([])
plt.yticks([])
plt.title('{}, start point_{}, goal viewpoint {}, {}\n dist = {:.4f} meter, theta = {:.4f} degree\n'.format(scene_name, point_idx, right_img_name[10:], str_succ, dist, theta_change))
plt.savefig('{}/{}'.format(point_folder, img_name), bbox_inches='tight', dpi=(400))
plt.close()
## ======================================================================================================================
## save stats
current_test_dict = {}
current_test_dict['img_name'] = right_img_name
current_test_dict['success_flag'] = flag_correct
current_test_dict['dist'] = dist
current_test_dict['theta'] = theta_change
current_test_dict['steps'] = count_steps
current_test_dict['collision'] = flag_broken
list_whole_stat.append(current_test_dict)
np.save('{}/runs_statistics.npy'.format(point_folder), list_whole_stat)
success_rate = 1.0 * count_correct / (len(point_pose_npy_file)-1)
print('count_correct/num_right_images = {}/{} = {}'.format(count_correct, len(point_pose_npy_file)-1, success_rate))
## write correctly run target image names to file
f = open('{}/successful_runs.txt'.format(point_folder), 'w')
f.write('count_correct/num_right_images = {}/{} = {}\n'.format(count_correct, len(point_pose_npy_file)-1, success_rate))
for i in range(len(list_correct_img_names)):
f.write('{}\n'.format(list_correct_img_names[i]))
f.close()
print('writing correct run image names to txt ...')
'''
avg_steps = 1.0 * sum(list_steps) / len(list_steps)
f.write('point {} : {}/{}, {}\n'.format(point_idx, count_correct, len(point_pose_npy_file)-1, avg_steps))
list_count_correct.append(count_correct)
list_count_runs.append(len(point_pose_npy_file)-1)
list_count_steps.append(avg_steps)
f.flush()
avg_count_steps = 1.0 * sum(list_count_steps) / len(list_count_steps)
f.write('In total : {}/{}, {}\n'.format(sum(list_count_correct), sum(list_count_runs), avg_count_steps))
f.write('-------------------------------------------------------------------------------------\n')
def main(scene_idx=0, point_a=0, right_a=0):
#scene_idx = 0
Train_Scenes, Test_Scenes = get_train_test_scenes()
mapper_scene2points = get_mapper_scene2points()
scene_name = Test_Scenes[scene_idx]
num_startPoints = len(mapper_scene2points[scene_name])
## as the move forward distance = 0.1, assume velocity is 0.01, it needs 10 steps.
def pose_interpolation(start_pose,
end_pose,
num=10,
include_endpoint=False):
x0, y0, theta0 = start_pose
x1, y1, theta1 = end_pose
x = np.linspace(x0, x1, num=num, endpoint=include_endpoint)
y = np.linspace(y0, y1, num=num, endpoint=include_endpoint)
## convert to quaternion
q0 = Quaternion(axis=[0, -1, 0], angle=theta0)
q1 = Quaternion(axis=[0, -1, 0], angle=theta1)
pose_list = []
v = np.array([1, 0, 0])
for idx, q in enumerate(
Quaternion.intermediates(q0,
q1,
num - 1,
include_endpoints=True)):
if idx < num:
e, d, f = q.rotate(v)
theta = atan2(f, e)
current_pose = [x[idx], y[idx], theta]
pose_list.append(current_pose)
return pose_list
## rrt functions
## first figure out how to sample points from rrt graph
rrt_directory = '/home/reza/Datasets/GibsonEnv/gibson/assets/dataset/{}_for_rrt'.format(
scene_name)
path_finder = rrt.PathFinder(rrt_directory)
path_finder.load()
num_nodes = len(path_finder.nodes_x)
free = cv2.imread(
'/home/reza/Datasets/GibsonEnv/gibson/assets/dataset/{}_for_rrt/free.png'
.format(scene_name), 0)
## draw the observations
## setup environment
import gym, logging
from mpi4py import MPI
from gibson.envs.husky_env import HuskyNavigateEnv
from baselines import logger
import skimage.io
from transforms3d.euler import euler2quat
config_file = os.path.join(
'/home/reza/Datasets/GibsonEnv/my_code/CVPR_workshop', 'env_yamls',
'{}_navigate.yaml'.format(scene_name))
env = HuskyNavigateEnv(config=config_file, gpu_count=1)
obs = env.reset(
) ## this line is important otherwise there will be an error like 'AttributeError: 'HuskyNavigateEnv' object has no attribute 'potential''
mapper_scene2z = get_mapper()
def get_obs(current_pose):
pos, orn = func_pose2posAndorn(current_pose,
mapper_scene2z[scene_name])
env.robot.reset_new_pose(pos, orn)
obs, _, _, _ = env.step(4)
obs_rgb = obs['rgb_filled']
obs_depth = obs['depth']
#obs_normal = obs['normal']
return obs_rgb, obs_depth #, obs_normal
def close_to_goal(pose1, pose2, thresh=0.20):
L2_dist = math.sqrt((pose1[0] - pose2[0])**2 +
(pose1[1] - pose2[1])**2)
thresh_L2_dist = thresh
theta_change = abs(pose1[2] - pose2[2]) / math.pi * 180
return (L2_dist <= thresh_L2_dist) #and (theta_change <= 30)
base_folder = '/home/reza/Datasets/GibsonEnv/my_code/visual_servoing/sample_image_pairs_{}'.format(
'test')
#for point_idx in range(2, 3):
for point_idx in range(point_a, point_a + 1):
print('point_idx = {}'.format(point_idx))
## read in start img and start pose
point_image_folder = '{}/{}/point_{}'.format(base_folder, scene_name,
point_idx)
point_pose_npy_file = np.load('{}/{}/point_{}_poses.npy'.format(
base_folder, scene_name, point_idx))
save_folder = '{}/{}/point_{}'.format(
'/home/reza/Datasets/GibsonEnv/my_code/vs_controller/for_video',
scene_name, point_idx)
## index 0 is the left image, so right_img_idx starts from index 1
#for right_img_idx in range(1, len(point_pose_npy_file)):
#for right_img_idx in range(1, 101):
for right_img_idx in range(right_a, right_a + 1):
right_img_name = point_pose_npy_file[right_img_idx]['img_name']
## Read in pose npy file generated from DQN
dqn_pose_npy_file = np.load(
'{}/run_{}/{}_waypoint_pose_list.npy'.format(
save_folder, right_img_idx, right_img_name[10:]))
start_pose = dqn_pose_npy_file[0]
goal_pose = dqn_pose_npy_file[1]
dqn_pose_list = dqn_pose_npy_file[2]
goal_img, goal_depth = get_obs(goal_pose)
goal_img = goal_img[:, :, ::-1]
cv2.imwrite(
'{}/run_{}/goal_img.jpg'.format(save_folder, right_img_idx),
goal_img)
interpolated_pose_list = []
## build the subsequence
len_dqn_pose_list = len(dqn_pose_list)
for i in range(len_dqn_pose_list - 1):
first_pose = dqn_pose_list[i]
second_pose = dqn_pose_list[i + 1]
subseq_pose_list = pose_interpolation(first_pose, second_pose)
interpolated_pose_list += subseq_pose_list
interpolated_pose_list.append(dqn_pose_list[-1])
img_name = 'goTo_{}.jpg'.format('current')
print('img_name = {}'.format(img_name))
## plot the poses
free2 = cv2.imread(
'/home/reza/Datasets/GibsonEnv/gibson/assets/dataset/{}_for_rrt/free.png'
.format(scene_name), 1)
rows, cols, _ = free2.shape
plt.imshow(free2)
for m in range(len(interpolated_pose_list)):
#for m in range(0, 100, 5):
pose = interpolated_pose_list[m]
x, y = path_finder.point_to_pixel((pose[0], pose[1]))
theta = pose[2]
plt.arrow(x, y, cos(theta), sin(theta), color='y', \
overhang=1, head_width=0.1, head_length=0.15, width=0.001)
## draw goal pose
x, y = path_finder.point_to_pixel((goal_pose[0], goal_pose[1]))
theta = goal_pose[2]
plt.arrow(x, y, cos(theta), sin(theta), color='r', \
overhang=1, head_width=0.1, head_length=0.15, width=0.001)
## draw start pose
x, y = path_finder.point_to_pixel((start_pose[0], start_pose[1]))
theta = start_pose[2]
plt.arrow(x, y, cos(theta), sin(theta), color='b', \
overhang=1, head_width=0.1, head_length=0.15, width=0.001)
plt.axis([0, cols, 0, rows])
plt.xticks([])
plt.yticks([])
#plt.title('{}, start point_{}, goal viewpoint {}, {}'.format(scene_name, point_idx, right_img_name[10:], str_succ))
plt.savefig('{}/run_{}/{}'.format(save_folder, right_img_idx,
'overview.jpg'),
bbox_inches='tight',
dpi=(400))
plt.close()
#'''
for i in range(len(interpolated_pose_list)):
current_pose = interpolated_pose_list[i]
obs_rgb, obs_depth = get_obs(current_pose)
obs_rgb = obs_rgb[:, :, ::-1]
cv2.imwrite(
'{}/run_{}/step_{}.jpg'.format(save_folder, right_img_idx, i),
obs_rgb)
sys.path.append('/home/reza/Datasets/GibsonEnv/my_code/CVPR_workshop')
sys.path.append('/home/reza/Datasets/GibsonEnv/my_code/visual_servoing')
sys.path.append('/home/reza/Datasets/GibsonEnv/my_code/DDPG')
from util_visual_servoing import get_train_test_scenes, get_mapper, get_mapper_scene2points, create_folder, get_mapper_dist_theta_heading, get_pose_from_name, sample_gt_random_dense_correspondences, sample_gt_dense_correspondences
import rrt
import random
from util import action2pose, func_pose2posAndorn, similar_location_under_certainThreshold, plus_theta_fn, minus_theta_fn
from util_vscontroller import gt_goToPose, genOverlapAreaOnCurrentView, genOverlapAreaOnGoalView, genErrorOverlapAreaOnCurrentView
from utils_ddpg import buildActionMapper, update_current_pose
from dqn_vs import DQN_vs_triplet
import torch
np.set_printoptions(precision=2, suppress=True)
## necessary constants
mapper_scene2points = get_mapper_scene2points()
num_episodes = 200000
batch_size = 128
lambda_action = 0.25
action_table = buildActionMapper(flag_fewer_actions=True)
seq_len = 50
def main(scene_idx=0, actual_episodes=1):
#scene_idx = 0
#actual_episodes=2
Train_Scenes, Test_Scenes = get_train_test_scenes()
scene_name = Train_Scenes[scene_idx]
num_startPoints = len(mapper_scene2points[scene_name])
model_weights_save_path = '{}'.format('/home/reza/Datasets/GibsonEnv/my_code/vs_controller/trained_dqn')
def main(scene_idx=0, point_a=0):
#scene_idx = 1
## necessary constants
mapper_scene2points = get_mapper_scene2points()
num_episodes = 200000
batch_size = 64
lambda_action = 0.25
action_table = buildActionMapper(flag_fewer_actions=True)
seq_len = 50
Train_Scenes, Test_Scenes = get_train_test_scenes()
if mode == 'Test':
scene_name = Test_Scenes[scene_idx]
elif mode == 'Train':
scene_name = Train_Scenes[scene_idx]
num_startPoints = len(mapper_scene2points[scene_name])
model_weights_save_path = '{}/{}'.format('/home/reza/Datasets/GibsonEnv/my_code/vs_controller/trained_dqn', approach)
action_space = action_table.shape[0]
##=============================================================================================================
## rrt functions
## first figure out how to sample points from rrt graph
rrt_directory = '/home/reza/Datasets/GibsonEnv/gibson/assets/dataset/{}_for_rrt'.format(scene_name)
path_finder = rrt.PathFinder(rrt_directory)
path_finder.load()
num_nodes = len(path_finder.nodes_x)
free = cv2.imread('/home/reza/Datasets/GibsonEnv/gibson/assets/dataset/{}_for_rrt/free.png'.format(scene_name), 0)
##------------------------------------------------------------------------------------------------------------
## setup environment
import gym, logging
from mpi4py import MPI
from gibson.envs.husky_env import HuskyNavigateEnv
from baselines import logger
import skimage.io
from transforms3d.euler import euler2quat
config_file = os.path.join('/home/reza/Datasets/GibsonEnv/my_code/CVPR_workshop', 'env_yamls', '{}_navigate.yaml'.format(scene_name))
env = HuskyNavigateEnv(config=config_file, gpu_count = 1)
obs = env.reset() ## this line is important otherwise there will be an error like 'AttributeError: 'HuskyNavigateEnv' object has no attribute 'potential''
mapper_scene2z = get_mapper()
def get_obs(current_pose):
pos, orn = func_pose2posAndorn(current_pose, mapper_scene2z[scene_name])
env.robot.reset_new_pose(pos, orn)
obs, _, _, _ = env.step(4)
obs_rgb = obs['rgb_filled']
obs_depth = obs['depth']
#obs_normal = obs['normal']
return obs_rgb, obs_depth#, obs_normal
def close_to_goal(pose1, pose2, thresh=0.20):
L2_dist = math.sqrt((pose1[0] - pose2[0])**2 + (pose1[1] - pose2[1])**2)
thresh_L2_dist = thresh
theta_change = abs(pose1[2] - pose2[2])/math.pi * 180
return (L2_dist <= thresh_L2_dist) #and (theta_change <= 30)
##============================================================================================================
if mode == 'Test':
base_folder = '/home/reza/Datasets/GibsonEnv/my_code/visual_servoing/sample_image_pairs_{}'.format('test')
elif mode == 'Train':
base_folder = '/home/reza/Datasets/GibsonEnv/my_code/visual_servoing/sample_image_pairs_{}'.format('train')
import torch
import torch.nn as nn
import torch.nn.functional as F
device = torch.device('cuda:0') ## Default CUDA device
num_epochs = 200000 ## same as # of trajs sampled
num_actions = action_table.shape[0]
if input_type == 'both':
perception = Perception_overlap(4).to(device)
elif input_type == 'siamese':
perception = Perception_siamese(4).to(device)
elif input_type == 'optical_flow':
perception = Perception_overlap(2).to(device)
elif input_type == 'optical_flow_depth':
perception = Perception_overlap(3).to(device)
elif input_type == 'optical_flow_depth_normalized':
perception = Perception_overlap(3).to(device)
elif input_type == 'optical_flow_depth_unnormalized_mask':
perception = Perception_overlap(3).to(device)
elif input_type == 'optical_flow_depth_siamese':
perception = Perception_siamese_fusion_new(3).to(device)
elif input_type == 'optical_flow_memory':
perception = Preception_overlap_resnet(4).to(device)
else:
perception = Perception_overlap(2).to(device)
if input_type == 'siamese':
model = DQN_OVERLAP_Controller(perception, num_actions, input_size=512).to(device)
elif input_type == 'optical_flow_memory':
model = DQN_OVERLAP_RESNET_Controller(perception, num_actions, input_size=512).to(device)
else:
model = DQN_OVERLAP_Controller(perception, num_actions, input_size=256).to(device)
model.load_state_dict(torch.load('{}/dqn_epoch_{}_Uvalda.pt'.format(model_weights_save_path, num_epochs)))
#model.eval()
list_succ = []
list_collision = []
## go through each point folder
if mode == 'Test':
a, b = 0, 1
elif mode == 'Train':
a, b = 7, 8
#a, b = 0, 1
#for point_idx in range(0, num_startPoints):
#for point_idx in range(a, b):
for point_idx in range(point_a, point_a+1):
print('point_idx = {}'.format(point_idx))
task_folder = '{}/{}/point_{}'.format('/home/reza/Datasets/GibsonEnv/my_code/vs_controller/for_video', scene_name, point_idx)
create_folder(task_folder)
## read in start img and start pose
point_image_folder = '{}/{}/point_{}'.format(base_folder, scene_name, point_idx)
point_pose_npy_file = np.load('{}/{}/point_{}_poses.npy'.format(base_folder, scene_name, point_idx))
#start_img = cv2.imread('{}/{}.png'.format(point_image_folder, point_pose_npy_file[0]['img_name']))[:, :, ::-1]
start_pose = point_pose_npy_file[0]['pose']
start_img, start_depth = get_obs(start_pose)
start_depth = start_depth.copy()
count_succ = 0
count_collision = 0
count_short_runs = 0
count_short_runs_collision = 0
count_short_runs_succ = 0
## index 0 is the left image, so right_img_idx starts from index 1
#for right_img_idx in range(1, len(point_pose_npy_file)):
for right_img_idx in range(1, 101):
print('right_img_idx = {}'.format(right_img_idx))
run_folder = '{}/run_{}'.format(task_folder, right_img_idx)
create_folder(run_folder)
current_pose = start_pose
right_img_name = point_pose_npy_file[right_img_idx]['img_name']
goal_pose = point_pose_npy_file[right_img_idx]['pose']
#goal_img = cv2.imread('{}/{}.png'.format(point_image_folder, right_img_name), 1)[:,:,::-1]
goal_img, goal_depth = get_obs(goal_pose)
goal_depth = goal_depth.copy()
current_depth = start_depth.copy()
episode_reward = 0
flag_succ = False
poses_list = []
poses_list.append(start_pose)
poses_list.append(goal_pose)
poses_list.append([current_pose])
for i_step in range(seq_len):
if input_type == 'both' or input_type == 'siamese':
overlapArea_currentView = genOverlapAreaOnCurrentView(current_depth, goal_depth, current_pose, goal_pose)[:,:,:2]
overlapArea_goalView = genOverlapAreaOnGoalView(current_depth, goal_depth, current_pose, goal_pose)[:,:,:2]
overlapArea = np.concatenate((overlapArea_currentView, overlapArea_goalView), axis=2)
elif input_type == 'optical_flow':
overlapArea = genGtDenseCorrespondenseFlowMap(current_depth, goal_depth, current_pose, goal_pose)[:,:,:2]
overlapArea = removeCorrespondenceRandomly(overlapArea, keep_prob=1.0)
elif input_type == 'optical_flow_depth':
opticalFlow = genGtDenseCorrespondenseFlowMap(current_depth, goal_depth, current_pose, goal_pose)[:,:,:2]
overlapArea = np.concatenate((opticalFlow, current_depth), axis=2)
elif input_type == 'optical_flow_depth_normalized':
opticalFlow = genGtDenseCorrespondenseFlowMap(current_depth, goal_depth, current_pose, goal_pose)[:,:,:2]
normalized_opticalFlow = normalize_opticalFlow(opticalFlow)
normalized_depth = normalize_depth(current_depth)
#normalized_depth = np.ones((256, 256, 1), np.float32)
overlapArea = np.concatenate((normalized_opticalFlow, normalized_depth), axis=2)
elif input_type == 'optical_flow_depth_unnormalized_mask':
opticalFlow, mask_flow = genGtDenseCorrespondenseFlowMapAndMask(current_depth, goal_depth, current_pose, goal_pose)
opticalFlow = opticalFlow[:, :, :2]
normalized_depth = current_depth * mask_flow
#normalized_opticalFlow = normalize_opticalFlow(opticalFlow)
normalized_depth = normalize_depth(normalized_depth)
overlapArea = np.concatenate((opticalFlow, normalized_depth), axis=2)
elif input_type == 'optical_flow_depth_siamese':
opticalFlow = genGtDenseCorrespondenseFlowMap(current_depth, goal_depth, current_pose, goal_pose)[:,:,:2]
normalized_depth = normalize_depth(current_depth)
#normalized_depth = np.ones((256, 256, 1), np.float32)
overlapArea = np.concatenate((opticalFlow, normalized_depth), axis=2)
#print('overlapArea.shape = {}'.format(overlapArea.shape))
elif input_type == 'optical_flow_memory':
opticalFlow = genGtDenseCorrespondenseFlowMap(current_depth, goal_depth, current_pose, goal_pose)[:,:,:2]
if i_step == 0:
overlapArea = np.concatenate((opticalFlow, opticalFlow), axis=2)
else:
overlapArea = np.concatenate((old_opticalFlow, opticalFlow), axis=2)
else:
overlapArea = genOverlapAreaOnCurrentView(current_depth, goal_depth, current_pose, goal_pose)[:,:,:2]
tensor_left = torch.tensor(overlapArea, dtype=torch.float32).to(device).unsqueeze(0).permute(0, 3, 1, 2)
Qvalue_table = model(tensor_left)
pred = Qvalue_table.max(1)[1].view(1, 1).detach().cpu().numpy().item() ## batch_size x 3
#print('Qvalue_table: {}'.format(Qvalue_table))
#print('pred = {}'.format(pred))
## update current_pose
vz, omegay = action_table[pred]
#print('vz = {:.2f}, omegay = {:.2f}'.format(vz, omegay))
vx = 0.0
vx = vx * lambda_action
vz = vz * lambda_action
omegay = omegay * pi * lambda_action
#print('actual velocity = {:.2f}, {:.2f}, {:.2f}'.format(vx, vz, omegay))
previous_pose = current_pose
current_pose = update_current_pose(current_pose, vx, vz, omegay)
poses_list[2].append(current_pose)
flag_broken = False
left_pixel = path_finder.point_to_pixel((previous_pose[0], previous_pose[1]))
right_pixel = path_finder.point_to_pixel((current_pose[0], current_pose[1]))
## rrt.line_check returns True when there is no obstacle
if not rrt.line_check(left_pixel, right_pixel, free):
print('run into something')
flag_broken = True
break
if close_to_goal(current_pose, goal_pose):
print('success run')
flag_succ = True
break
## compute new_state
current_img, current_depth = get_obs(current_pose)
current_depth = current_depth.copy()
#old_opticalFlow = opticalFlow.copy()
np.save('{}/{}_waypoint_pose_list.npy'.format(run_folder, right_img_name[10:]), poses_list)
#assert 1==2
if flag_succ:
count_succ += 1
list_succ.append(point_pose_npy_file[right_img_idx]['img_name'])
if findShortRangeImageName(right_img_name):
count_short_runs_succ += 1
if flag_broken:
count_collision += 1
list_collision.append(point_pose_npy_file[right_img_idx]['img_name'])
if findShortRangeImageName(right_img_name):
count_short_runs_collision += 1
if findShortRangeImageName(right_img_name):
count_short_runs += 1
print('count_succ = {}'.format(count_succ))
print('count_collision = {}'.format(count_collision))
print('count_short_runs_succ = {}'.format(count_short_runs_succ))
print('count_short_runs_collision = {}'.format(count_short_runs_collision))
print('num_succ = {}, num_run = {}, count_short_runs_succ = {}, count_short_runs = {}'.format(count_succ, len(point_pose_npy_file), count_short_runs_succ, count_short_runs))