OpenAI empêche à un ingénieur qui a créé un fusil robotisé piloté par ChatGPT d'utiliser son API,

L'entreprise participe pourtant à des projets d'armement piloté par l'intelligence artificielle

OpenAI empêche à un ingénieur qui a créé un fusil robotisé piloté par ChatGPT d'utiliser son API
L’entreprise participe pourtant à des projets d'armement piloté par l'intelligence artificielle

OpenAI empêche un développeur qui a mis sur pied un fusil robotisé piloté par ChatGPT d’utiliser son API. L’entreprise a pourtant supprimé l’interdiction de faire usage de ChatGPT à des fins militaires de ses conditions d’utilisation. Elle participe d’ailleurs à des projets militaires qui mettent l’intelligence artificielle à contribution sur les champs de bataille et en cela alimente la course à l’armement piloté par l’IA à l’échelle mondiale.

Guy on tiktok (@sts_3d) used the #OpenAI realtime api to make a voice controlled smart gun pic.twitter.com/VGGxBkDEv9

— andrew gao (@itsandrewgao) January 7, 2025

« ChatGPT, nous sommes attaqués de l'avant gauche et de l'avant droite », dit le développeur au système dans la vidéo. « Réagissez en conséquence. La vitesse et la précision avec lesquelles le fusil répond sont impressionnantes, car il s'appuie sur l'API en temps réel d'OpenAI pour interpréter les données et renvoyer des instructions que l'engin peut comprendre. Il suffirait d'une simple formation pour que ChatGPT reçoive une commande telle que « tourne à gauche » et comprenne comment la traduire dans un langage lisible par la machine. Dans une récente déclaration, OpenAI a indiqué avoir sevré le développeur qui en était à l'origine de la possibilité de faire usage de son API.

OpenAI venait pourtant d’ajouter Anduril à sa liste de partenaires dans le cadre de l’utilisation de l’intelligence artificielle à des fins militaires

« Nous joignons nos forces à celles d'@OpenAI pour faire progresser les solutions d'IA pour la sécurité nationale. L'Amérique doit gagner. Les modèles d'OpenAI combinés aux systèmes de défense d'Anduril protégeront le personnel militaire américain et allié des attaques de drones et amélioreront la prise de décision en temps réel. Dans la course mondiale à l'IA, ce partenariat témoigne de notre engagement commun à faire en sorte que les forces américaines et alliées aient accès aux technologies d'intelligence artificielle les plus avancées et les plus responsables au monde », annonce Anduril qui parle sans détour de course à l’armement.

Les deux entreprises travailleront ensemble à l'intégration de la technologie d'OpenAI dans le logiciel des systèmes anti-drones d'Anduril.

« OpenAI développe l'intelligence artificielle pour qu'elle profite au plus grand nombre et soutient les efforts déployés par les États-Unis pour veiller à ce que la technologie respecte les valeurs démocratiques. Notre partenariat avec Anduril contribuera à garantir que la technologie OpenAI protège le personnel militaire américain et aidera la communauté de la sécurité nationale à comprendre et à utiliser de manière responsable cette technologie pour assurer la sécurité et la liberté de nos concitoyens », déclare Sam Altman.

We’re joining forces with @OpenAI to advance AI solutions for national security.

America needs to win.

OpenAI’s models combined with Anduril’s defense systems will protect U.S. and allied military personnel from attacks by unmanned drones and improve real-time decision-making.… pic.twitter.com/fprGN05eeg

— Anduril Industries (@anduriltech) December 4, 2024

OpenAI alimente ainsi la course à l’armement piloté par l’intelligence artificielle à laquelle on assiste à l’échelle mondiale

L'armée chinoise a dévoilé un nouveau type de compagnon de combat pour ses soldats : un robot-chien avec une mitrailleuse attachée à son dos. Dans une vidéo diffusée par l'agence de presse gouvernementale CCTV, on voit des membres des militaires chinois opérer sur un champ de tir aux côtés d'un robot à quatre pattes sur lequel est monté ce qui semble être une variante du fusil d'assaut QBZ-95 de 5,8 x 42 mm, dans le cadre des récents exercices militaires conjoints Golden Dragon 24 menés par la Chine et le Cambodge dans le golfe de Thaïlande.

Ces robots s’appuient à la base sur des applications de détection et suivi d’objets. Dans ce cas, il y a au préalable collecte des images provenant de caméras avant puis détection d’objet sur une classe spécifiée. Cette détection utilise Tensorflow via le tensorflow_object_detector. Le robot accepte n'importe quel modèle Tensorflow et permet au développeur de spécifier un sous-ensemble de classes de détection incluses dans le modèle. Il effectue cet ensemble d'opérations pour un nombre prédéfini d'itérations, en bloquant pendant une durée prédéfinie entre chaque itération. L'application détermine ensuite l'emplacement de la détection la plus fiable de la classe spécifiée et se dirige vers l'objet.

L’application est organisée en trois ensembles de processus Python communiquant avec le robot Spot. Le diagramme des processus est illustré ci-dessous. Le processus principal communique avec le robot Spot via GRPC et reçoit constamment des images. Ces images sont poussées dans la RAW_IMAGES_QUEUE et lues par les processus Tensorflow. Ces processus détectent des objets dans les images et poussent l'emplacement dans PROCESSED_BOXES_QUEUE. Le thread principal détermine alors l'emplacement de l'objet et envoie des commandes au robot pour qu'il se dirige vers l'objet.

# Copyright (c) 2023 Boston Dynamics, Inc. All rights reserved.
#
# Downloading, reproducing, distributing or otherwise using the SDK Software
# is subject to the terms and conditions of the Boston Dynamics Software
# Development Kit License (20191101-BDSDK-SL).

"""Tutorial to show how to use the Boston Dynamics API to detect and follow an object"""
import argparse
import io
import json
import math
import os
import signal
import sys
import time
from multiprocessing import Barrier, Process, Queue, Value
from queue import Empty, Full
from threading import BrokenBarrierError, Thread

import cv2
import numpy as np
from PIL import Image
from scipy import ndimage
from tensorflow_object_detection import DetectorAPI

import bosdyn.client
import bosdyn.client.util
from bosdyn import geometry
from bosdyn.api import geometry_pb2 as geo
from bosdyn.api import image_pb2, trajectory_pb2
from bosdyn.api.image_pb2 import ImageSource
from bosdyn.api.spot import robot_command_pb2 as spot_command_pb2
from bosdyn.client.async_tasks import AsyncPeriodicQuery, AsyncTasks
from bosdyn.client.frame_helpers import (GROUND_PLANE_FRAME_NAME, VISION_FRAME_NAME, get_a_tform_b,
get_vision_tform_body)
from bosdyn.client.image import ImageClient
from bosdyn.client.lease import LeaseClient, LeaseKeepAlive
from bosdyn.client.math_helpers import Quat, SE3Pose
from bosdyn.client.robot_command import (CommandFailedError, CommandTimedOutError,
RobotCommandBuilder, RobotCommandClient, blocking_stand)
from bosdyn.client.robot_state import RobotStateClient

LOGGER = bosdyn.client.util.get_logger()

SHUTDOWN_FLAG = Value('i', 0)

# Don't let the queues get too backed up
QUEUE_MAXSIZE = 10

# This is a multiprocessing.Queue for communication between the main process and the
# Tensorflow processes.
# Entries in this queue are in the format:

# {
# 'source': Name of the camera,
# 'world_tform_cam': transform from VO to camera,
# 'world_tform_gpe': transform from VO to ground plane,
# 'raw_image_time': Time when the image was collected,
# 'cv_image': The decoded image,
# 'visual_dims': (cols, rows),
# 'depth_image': depth image proto,
# 'system_cap_time': Time when the image was received by the main process,
# 'image_queued_time': Time when the image was done preprocessing and queued
# }
RAW_IMAGES_QUEUE = Queue(QUEUE_MAXSIZE)

# This is a multiprocessing.Queue for communication between the Tensorflow processes and
# the bbox print process. This is meant for running in a containerized environment with no access
# to an X display
# Entries in this queue have the following fields in addition to those in :
# {
# 'processed_image_start_time': Time when the image was received by the TF process,
# 'processed_image_end_time': Time when the image was processing for bounding boxes
# 'boxes': list of detected bounding boxes for the processed image
# 'classes': classes of objects,
# 'scores': confidence scores,
# }
PROCESSED_BOXES_QUEUE = Queue(QUEUE_MAXSIZE)

# Barrier for waiting on Tensorflow processes to start, initialized in main()
TENSORFLOW_PROCESS_BARRIER = None

COCO_CLASS_DICT = {
1: 'person',
2: 'bicycle',
3: 'car',
4: 'motorcycle',
5: 'airplane',
6: 'bus',
7: 'train',
8: 'truck',
9: 'boat',
10: 'trafficlight',
11: 'firehydrant',
13: 'stopsign',
14: 'parkingmeter',
15: 'bench',
16: 'bird',
17: 'cat',
18: 'dog',
19: 'horse',
20: 'sheep',
21: 'cow',
22: 'elephant',
23: 'bear',
24: 'zebra',
25: 'giraffe',
27: 'backpack',
28: 'umbrella',
31: 'handbag',
32: 'tie',
33: 'suitcase',
34: 'frisbee',
35: 'skis',
36: 'snowboard',
37: 'sportsball',
38: 'kite',
39: 'baseballbat',
40: 'baseballglove',
41: 'skateboard',
42: 'surfboard',
43: 'tennisracket',
44: 'bottle',
46: 'wineglass',
47: 'cup',
48: 'fork',
49: 'knife',
50: 'spoon',
51: 'bowl',
52: 'banana',
53: 'apple',
54: 'sandwich',
55: 'orange',
56: 'broccoli',
57: 'carrot',
58: 'hotdog',
59: 'pizza',
60: 'donut',
61: 'cake',
62: 'chair',
63: 'couch',
64: 'pottedplant',
65: 'bed',
67: 'diningtable',
70: 'toilet',
72: 'tv',
73: 'laptop',
74: 'mouse',
75: 'remote',
76: 'keyboard',
77: 'cellphone',
78: 'microwave',
79: 'oven',
80: 'toaster',
81: 'sink',
82: 'refrigerator',
84: 'book',
85: 'clock',
86: 'vase',
87: 'scissors',
88: 'teddybear',
89: 'hairdrier',
90: 'toothbrush'
}

# Mapping from visual to depth data
VISUAL_SOURCE_TO_DEPTH_MAP_SOURCE = {
'frontleft_fisheye_image': 'frontleft_depth_in_visual_frame',
'frontright_fisheye_image': 'frontright_depth_in_visual_frame'
}
ROTATION_ANGLES = {
'back_fisheye_image': 0,
'frontleft_fisheye_image': -78,
'frontright_fisheye_image': -102,
'left_fisheye_image': 0,
'right_fisheye_image': 180
}


def _update_thread(async_task):
while True:
async_task.update()
time.sleep(0.01)


class AsyncImage(AsyncPeriodicQuery):
"""Grab image."""

def __init__(self, image_client, image_sources):
# Period is set to be about 15 FPS
super(AsyncImage, self).__init__('images', image_client, LOGGER, period_sec=0.067)
self.image_sources = image_sources

def _start_query(self):
return self._client.get_image_from_sources_async(self.image_sources)


class AsyncRobotState(AsyncPeriodicQuery):
"""Grab robot state."""

def __init__(self, robot_state_client):
# period is set to be about the same rate as detections on the CORE AI
super(AsyncRobotState, self).__init__('robot_state', robot_state_client, LOGGER,
period_sec=0.02)

def _start_query(self):
return self._client.get_robot_state_async()


def get_source_list(image_client):
"""Gets a list of image sources and filters based on config dictionary

Args:
image_client: Instantiated image client
"""

# We are using only the visual images with their corresponding depth sensors
sources = image_client.list_image_sources()
source_list = []
for source in sources:
if source.image_type == ImageSource.IMAGE_TYPE_VISUAL:
# only append if sensor has corresponding depth sensor
if source.name in VISUAL_SOURCE_TO_DEPTH_MAP_SOURCE:
source_list.append(source.name)
source_list.append(VISUAL_SOURCE_TO_DEPTH_MAP_SOURCE[source.name])
return source_list


def capture_images(image_task, sleep_between_capture):
""" Captures images and places them on the queue

Args:
image_task (AsyncImage): Async task that provides the images response to use
sleep_between_capture (float): Time to sleep between each image capture
"""
while not SHUTDOWN_FLAG.value:
get_im_resp = image_task.proto
start_time = time.time()
if not get_im_resp:
continue
depth_responses = {
img.source.name: img
for img in get_im_resp
if img.source.image_type == ImageSource.IMAGE_TYPE_DEPTH
}
entry = {}
for im_resp in get_im_resp:
if im_resp.source.image_type == ImageSource.IMAGE_TYPE_VISUAL:
source = im_resp.source.name
depth_source = VISUAL_SOURCE_TO_DEPTH_MAP_SOURCE
depth_image = depth_responses

acquisition_time = im_resp.shot.acquisition_time
image_time = acquisition_time.seconds + acquisition_time.nanos * 1e-9

try:
image = Image.open(io.BytesIO(im_resp.shot.image.data))
source = im_resp.source.name

image = ndimage.rotate(image, ROTATION_ANGLES)
if im_resp.shot.image.pixel_format == image_pb2.Image.PIXEL_FORMAT_GREYSCALE_U8:
image = cv2.cvtColor(image, cv2.COLOR_GRAY2RGB) # Converted to RGB for TF
tform_snapshot = im_resp.shot.transforms_snapshot
frame_name = im_resp.shot.frame_name_image_sensor
world_tform_cam = get_a_tform_b(tform_snapshot, VISION_FRAME_NAME, frame_name)
world_tform_gpe = get_a_tform_b(tform_snapshot, VISION_FRAME_NAME,
GROUND_PLANE_FRAME_NAME)
entry = {
'source': source,
'world_tform_cam': world_tform_cam,
'world_tform_gpe': world_tform_gpe,
'raw_image_time': image_time,
'cv_image': image,
'visual_dims': (im_resp.shot.image.cols, im_resp.shot.image.rows),
'depth_image': depth_image,
'system_cap_time': start_time,
'image_queued_time': time.time()
}
except Exception as exc: # pylint: disable=broad-except
print(f'Exception occurred during image capture {exc}')
try:
RAW_IMAGES_QUEUE.put_nowait(entry)
except Full as exc:
print(f'RAW_IMAGES_QUEUE is full: {exc}')
time.sleep(sleep_between_capture)


def start_tensorflow_processes(num_processes, model_path, detection_class, detection_threshold,
max_processing_delay):
"""Starts Tensorflow processes in parallel.

It does not keep track of the processes once they are started because they run indefinitely
and are never joined back to the main process.

Args:
num_processes (int): Number of Tensorflow processes to start in parallel.
model_path (str): Filepath to the Tensorflow model to use.
detection_class (int): Detection class to detect
detection_threshold (float): Detection threshold to apply to all Tensorflow detections.
max_processing_delay (float): Allowed delay before processing an incoming image.
"""
processes = []
for _ in range(num_processes):
process = Process(
target=process_images, args=(
model_path,
detection_class,
detection_threshold,
max_processing_delay,
), daemon=True)
process.start()
processes.append(process)
return processes


def process_images(model_path, detection_class, detection_threshold, max_processing_delay):
"""Starts Tensorflow and detects objects in the incoming images.

Args:
model_path (str): Filepath to the Tensorflow model to use.
detection_class (int): Detection class to detect
detection_threshold (float): Detection threshold to apply to all Tensorflow detections.
max_processing_delay (float): Allowed delay before processing an incoming image.
"""

odapi = DetectorAPI(path_to_ckpt=model_path)
num_processed_skips = 0

if TENSORFLOW_PROCESS_BARRIER is None:
return

try:
TENSORFLOW_PROCESS_BARRIER.wait()
except BrokenBarrierError as exc:
print(f'Error waiting for Tensorflow processes to initialize: {exc}')
return False

while not SHUTDOWN_FLAG.value:
try:
entry = RAW_IMAGES_QUEUE.get_nowait()
except Empty:
time.sleep(0.1)
continue
for _, capture in entry.items():
start_time = time.time()
processing_delay = time.time() - capture['raw_image_time']
if processing_delay > max_processing_delay:
num_processed_skips += 1
print(f'skipped image because it took {processing_delay}')
continue # Skip image due to delay

image = capture['cv_image']
boxes, scores, classes, _ = odapi.process_frame(image)
confident_boxes = []
confident_object_classes = []
confident_scores = []
if len(boxes) == 0:
print('no detections founds')
continue
for box, score, box_class in sorted(zip(boxes, scores, classes), key=lambda x: x,
reverse=True):
if score > detection_threshold and box_class == detection_class:
confident_boxes.append(box)
confident_object_classes.append(COCO_CLASS_DICT)
confident_scores.append(score)
image = cv2.rectangle(image, (box, box), (box, box), (255, 0, 0), 2)

capture['processed_image_start_time'] = start_time
capture['processed_image_end_time'] = time.time()
capture['boxes'] = confident_boxes
capture['classes'] = confident_object_classes
capture['scores'] = confident_scores
capture['cv_image'] = image
try:
PROCESSED_BOXES_QUEUE.put_nowait(entry)
except Full as exc:
print(f'PROCESSED_BOXES_QUEUE is full: {exc}')
print('tf process ending')
return True


def get_go_to(world_tform_object, robot_state, mobility_params, dist_margin=0.5):
"""Gets trajectory command to a goal location

Args:
world_tform_object (SE3Pose): Transform from vision frame to target object
robot_state (RobotState): Current robot state
mobility_params (MobilityParams): Mobility parameters
dist_margin (float): Distance margin to target
"""
vo_tform_robot = get_vision_tform_body(robot_state.kinematic_state.transforms_snapshot)
print(f'robot pos: {vo_tform_robot}')
delta_ewrt_vo = np.array(
[world_tform_object.x - vo_tform_robot.x, world_tform_object.y - vo_tform_robot.y, 0])
norm = np.linalg.norm(delta_ewrt_vo)
if norm == 0:
return None
delta_ewrt_vo_norm = delta_ewrt_vo / norm
heading = _get_heading(delta_ewrt_vo_norm)
vo_tform_goal = np.array([
world_tform_object.x - delta_ewrt_vo_norm * dist_margin,
world_tform_object.y - delta_ewrt_vo_norm * dist_margin
])
se2_pose = geo.SE2Pose(position=geo.Vec2(x=vo_tform_goal, y=vo_tform_goal), angle=heading)
tag_cmd = RobotCommandBuilder.synchro_se2_trajectory_command(se2_pose,
frame_name=VISION_FRAME_NAME,
params=mobility_params)
return tag_cmd


def _get_heading(xhat):
zhat = [0.0, 0.0, 1.0]
yhat = np.cross(zhat, xhat)
mat = np.array([xhat, yhat, zhat]).transpose()
return Quat.from_matrix(mat).to_yaw()


def set_default_body_control():
"""Set default body control params to current body position"""
footprint_R_body = geometry.EulerZXY()
position = geo.Vec3(x=0.0, y=0.0, z=0.0)
rotation = footprint_R_body.to_quaternion()
pose = geo.SE3Pose(position=position, rotation=rotation)
point = trajectory_pb2.SE3TrajectoryPoint(pose=pose)
traj = trajectory_pb2.SE3Trajectory(points=)
return spot_command_pb2.BodyControlParams(base_offset_rt_footprint=traj)


def get_mobility_params():
"""Gets mobility parameters for following"""
vel_desired = .75
speed_limit = geo.SE2VelocityLimit(
max_vel=geo.SE2Velocity(linear=geo.Vec2(x=vel_desired, y=vel_desired), angular=.25))
body_control = set_default_body_control()
mobility_params = spot_command_pb2.MobilityParams(vel_limit=speed_limit, obstacle_params=None,
body_control=body_control,
locomotion_hint=spot_command_pb2.HINT_TROT)
return mobility_params


def depth_to_xyz(depth, pixel_x, pixel_y, focal_length, principal_point):
"""Calculate the transform to point in image using camera intrinsics and depth"""
x = depth * (pixel_x - principal_point.x) / focal_length.x
y = depth * (pixel_y - principal_point.y) / focal_length.y
z = depth
return x, y, z


def remove_ground_from_depth_image(raw_depth_image, focal_length, principal_point, world_tform_cam,
world_tform_gpe, ground_tolerance=0.04):
""" Simple ground plane removal algorithm. Uses ground height
and does simple z distance filtering.

Args:
raw_depth_image (np.array): Depth image
focal_length (Vec2): Focal length of camera that produced the depth image
principal_point (Vec2): Principal point of camera that produced the depth image
world_tform_cam (SE3Pose): Transform from VO to camera frame
world_tform_gpe (SE3Pose): Transform from VO to GPE frame
ground_tolerance (float): Distance in meters to add to the ground plane
"""
new_depth_image = raw_depth_image

# same functions as depth_to_xyz, but converted to np functions
indices = np.indices(raw_depth_image.shape)
xs = raw_depth_image * (indices - principal_point.x) / focal_length.x
ys = raw_depth_image * (indices - principal_point.y) / focal_length.y
zs = raw_depth_image

# create xyz point cloud
camera_tform_points = np.stack([xs, ys, zs], axis=2)
# points in VO frame
world_tform_points = world_tform_cam.transform_cloud(camera_tform_points)
# array of booleans where True means the point was below the ground plane plus tolerance
world_tform_points_mask = (world_tform_gpe.z - world_tform_points[:, :, 2]) < ground_tolerance
# remove data below ground plane
new_depth_image = 0
return new_depth_image


def get_distance_to_closest_object_depth(x_min, x_max, y_min, y_max, depth_scale, raw_depth_image,
histogram_bin_size=0.50, minimum_number_of_points=10,
max_distance=8.0):
"""Make a histogram of distances to points in the cloud and take the closest distance with
enough points.

Args:
x_min (int): minimum x coordinate (column) of object to find
x_max (int): maximum x coordinate (column) of object to find
y_min (int): minimum y coordinate (row) of object to find
y_max (int): maximum y coordinate (row) of object to find
depth_scale (float): depth scale of the image to convert from sensor value to meters
raw_depth_image (np.array): matrix of depth pixels
histogram_bin_size (float): size of each bin of distances
minimum_number_of_points (int): minimum number of points before returning depth
max_distance (float): maximum distance to object in meters
"""
num_bins = math.ceil(max_distance / histogram_bin_size)

# get a sub-rectangle of the bounding box out of the whole image, then flatten
obj_depths = (raw_depth_image[y_min:y_max, x_min:x_max]).flatten()
obj_depths = obj_depths / depth_scale
obj_depths = obj_depths[obj_depths != 0]

hist, hist_edges = np.histogram(obj_depths, bins=num_bins, range=(0, max_distance))

edges_zipped = zip(hist_edges[:-1], hist_edges[1:])
# Iterate over the histogram and return the first distance with enough points.
for entry, edges in zip(hist, edges_zipped):
if entry > minimum_number_of_points:
filtered_depths = obj_depths[(obj_depths > edges) & (obj_depths < edges)]
if len(filtered_depths) == 0:
continue
return np.mean(filtered_depths)

return max_distance


def rotate_about_origin_degrees(origin, point, angle):
"""
Rotate a point counterclockwise by a given angle around a given origin.

Args:
origin (tuple): Origin to rotate the point around
point (tuple): Point to rotate
angle (float): Angle in degrees
"""
return rotate_about_origin(origin, point, math.radians(angle))


def rotate_about_origin(origin, point, angle):
"""
Rotate a point counterclockwise by a given angle around a given origin.

Args:
origin (tuple): Origin to rotate the point around
point (tuple): Point to rotate
angle (float): Angle in radians
"""
orig_x, orig_y = origin
pnt_x, pnt_y = point

ret_x = orig_x + math.cos(angle) * (pnt_x - orig_x) - math.sin(angle) * (pnt_y - orig_y)
ret_y = orig_y + math.sin(angle) * (pnt_x - orig_x) + math.cos(angle) * (pnt_y - orig_y)
return int(ret_x), int(ret_y)


def get_object_position(world_tform_cam, world_tform_gpe, visual_dims, depth_image, bounding_box,
rotation_angle):
"""
Extract the bounding box, then find the mode in that region.

Args:
world_tform_cam (SE3Pose): SE3 transform from world to camera frame
visual_dims (Tuple): (cols, rows) tuple from the visual image
depth_image (ImageResponse): From a depth camera corresponding to the visual_image
bounding_box (list): Bounding box from tensorflow
rotation_angle (float): Angle (in degrees) to rotate depth image to match cam image rotation
"""

# Make sure there are two images.
if visual_dims is None or depth_image is None:
# Fail.
return

# Rotate bounding box back to original frame
points = [(bounding_box, bounding_box), (bounding_box, bounding_box),
(bounding_box, bounding_box), (bounding_box, bounding_box)]

origin = (visual_dims / 2, visual_dims / 2)

points_rot = [rotate_about_origin_degrees(origin, point, rotation_angle) for point in points]

# Get the bounding box corners.
y_min = max(0, min([point for point in points_rot]))
x_min = max(0, min([point for point in points_rot]))
y_max = min(visual_dims, max([point for point in points_rot]))
x_max = min(visual_dims, max([point for point in points_rot]))

# Check that the bounding box is valid.
if (x_min < 0 or y_min < 0 or x_max > visual_dims or y_max > visual_dims):
print(f'Bounding box is invalid: ({x_min}, {y_min}) | ({x_max}, {y_max})')
print(f'Bounds: ({visual_dims}, {visual_dims})')
return

# Unpack the images.
try:
if depth_image.shot.image.pixel_format == image_pb2.Image.PIXEL_FORMAT_DEPTH_U16:
dtype = np.uint16
else:
dtype = np.uint8
img = np.fromstring(depth_image.shot.image.data, dtype=dtype)
if depth_image.shot.image.format == image_pb2.Image.FORMAT_RAW:
img = img.reshape(depth_image.shot.image.rows, depth_image.shot.image.cols)
else:
img = cv2.imdecode(img, -1)
depth_image_pixels = img
depth_image_pixels = remove_ground_from_depth_image(
depth_image_pixels, depth_image.source.pinhole.intrinsics.focal_length,
depth_image.source.pinhole.intrinsics.principal_point, world_tform_cam, world_tform_gpe)
# Get the depth data from the region in the bounding box.
max_distance = 8.0
depth = get_distance_to_closest_object_depth(x_min, x_max, y_min, y_max,
depth_image.source.depth_scale,
depth_image_pixels, max_distance=max_distance)

if depth >= max_distance:
# Not enough depth data.
print('Not enough depth data.')
return False
else:
print(f'distance to object: {depth}')

center_x = round((x_max - x_min) / 2.0 + x_min)
center_y = round((y_max - y_min) / 2.0 + y_min)

tform_x, tform_y, tform_z = depth_to_xyz(
depth, center_x, center_y, depth_image.source.pinhole.intrinsics.focal_length,
depth_image.source.pinhole.intrinsics.principal_point)
camera_tform_obj = SE3Pose(tform_x, tform_y, tform_z, Quat())

return world_tform_cam * camera_tform_obj
except Exception as exc: # pylint: disable=broad-except
print(f'Error getting object position: {exc}')
return


def _check_model_path(model_path):
if model_path is None or \
not os.path.exists(model_path) or \
not os.path.isfile(model_path):
print(f'ERROR, could not find model file {model_path}')
return False
return True


def _check_and_load_json_classes(config_path):
if os.path.isfile(config_path):
with open(config_path) as json_classes:
global COCO_CLASS_DICT # pylint: disable=global-statement
COCO_CLASS_DICT = json.load(json_classes)


def _find_highest_conf_source(processed_boxes_entry):
highest_conf_source = None
max_score = 0
for key, capture in processed_boxes_entry.items():
if 'scores' in capture.keys():
if len(capture['scores']) > 0 and capture['scores'] > max_score:
highest_conf_source = key
max_score = capture['scores']
return highest_conf_source


def signal_handler(signal, frame):
print('Interrupt caught, shutting down')
SHUTDOWN_FLAG.value = 1


def main():
"""Command line interface."""

parser = argparse.ArgumentParser()
parser.add_argument(
'--model-path', default='/model.pb', help=
('Local file path to the Tensorflow model, example pre-trained models can be found at '
'https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/tf1_detection_zoo.md'
))
parser.add_argument('--classes', default='/classes.json', type=str,
help='File containing json mapping of object class IDs to class names')
parser.add_argument('--number-tensorflow-processes', default=1, type=int,
help='Number of Tensorflow processes to run in parallel')
parser.add_argument('--detection-threshold', default=0.7, type=float,
help='Detection threshold to use for Tensorflow detections')
parser.add_argument(
'--sleep-between-capture', default=0.2, type=float,
help=('Seconds to sleep between each image capture loop iteration, which captures '
'an image from all cameras'))
parser.add_argument(
'--detection-class', default=1, type=int,
help=('Detection classes to use in the Tensorflow model.'
'Default is to use 1, which is a person in the Coco dataset'))
parser.add_argument(
'--max-processing-delay', default=7.0, type=float,
help=('Maximum allowed delay for processing an image. '
'Any image older than this value will be skipped'))
parser.add_argument('--test-mode', action='store_true',
help='Run application in test mode, don\'t execute commands')

bosdyn.client.util.add_base_arguments(parser)
bosdyn.client.util.add_payload_credentials_arguments(parser)
options = parser.parse_args()
signal.signal(signal.SIGINT, signal_handler)
try:
# Make sure the model path is a valid file
if not _check_model_path(options.model_path):
return False

# Check for classes json file, otherwise use the COCO class dictionary
_check_and_load_json_classes(options.classes)

global TENSORFLOW_PROCESS_BARRIER # pylint: disable=global-statement
TENSORFLOW_PROCESS_BARRIER = Barrier(options.number_tensorflow_processes + 1)
# Start Tensorflow processes
tf_processes = start_tensorflow_processes(options.number_tensorflow_processes,
options.model_path, options.detection_class,
options.detection_threshold,
options.max_processing_delay)

# sleep to give the Tensorflow processes time to initialize
try:
TENSORFLOW_PROCESS_BARRIER.wait()
except BrokenBarrierError as exc:
print(f'Error waiting for Tensorflow processes to initialize: {exc}')
return False
# Start the API related things

# Create robot object with a world object client
sdk = bosdyn.client.create_standard_sdk('SpotFollowClient')
robot = sdk.create_robot(options.hostname)

if options.payload_credentials_file:
robot.authenticate_from_payload_credentials(
*bosdyn.client.util.get_guid_and_secret(options))
else:
bosdyn.client.util.authenticate(robot)

# Time sync is necessary so that time-based filter requests can be converted
robot.time_sync.wait_for_sync()

# Verify the robot is not estopped and that an external application has registered and holds
# an estop endpoint.
assert not robot.is_estopped(), 'Robot is estopped. Please use an external E-Stop client,' \
' such as the estop SDK example, to configure E-Stop.'

# Create the sdk clients
robot_state_client = robot.ensure_client(RobotStateClient.default_service_name)
robot_command_client = robot.ensure_client(RobotCommandClient.default_service_name)
lease_client = robot.ensure_client(LeaseClient.default_service_name)
image_client = robot.ensure_client(ImageClient.default_service_name)
source_list = get_source_list(image_client)
image_task = AsyncImage(image_client, source_list)
robot_state_task = AsyncRobotState(robot_state_client)
task_list = [image_task, robot_state_task]
_async_tasks = AsyncTasks(task_list)
print('Detect and follow client connected.')

lease = lease_client.take()
lease_keep = LeaseKeepAlive(lease_client)
# Power on the robot and stand it up
resp = robot.power_on()
try:
blocking_stand(robot_command_client)
except CommandFailedError as exc:
print(f'Error ({exc}) occurred while trying to stand. Check robot surroundings.')
return False
except CommandTimedOutError as exc:
print(f'Stand command timed out: {exc}')
return False
print('Robot powered on and standing.')
params_set = get_mobility_params()

# This thread starts the async tasks for image and robot state retrieval
update_thread = Thread(target=_update_thread, args=)
update_thread.daemon = True
update_thread.start()
# Wait for the first responses.
while any(task.proto is None for task in task_list):
time.sleep(0.1)

# Start image capture process
image_capture_thread = Process(target=capture_images,
args=(image_task, options.sleep_between_capture),
daemon=True)
image_capture_thread.start()
while not SHUTDOWN_FLAG.value:
# This comes from the tensorflow processes and limits the rate of this loop
try:
entry = PROCESSED_BOXES_QUEUE.get_nowait()
except Empty:
continue
# find the highest confidence bounding box
highest_conf_source = _find_highest_conf_source(entry)
if highest_conf_source is None:
# no boxes or scores found
continue
capture_to_use = entry
raw_time = capture_to_use['raw_image_time']
time_gap = time.time() - raw_time
if time_gap > options.max_processing_delay:
continue # Skip image due to delay

# Find the transform to the highest confidence object using the depth sensor
get_object_position_start = time.time()
robot_state = robot_state_task.proto
world_tform_gpe = get_a_tform_b(robot_state.kinematic_state.transforms_snapshot,
VISION_FRAME_NAME, GROUND_PLANE_FRAME_NAME)
world_tform_object = get_object_position(
capture_to_use['world_tform_cam'], world_tform_gpe, capture_to_use['visual_dims'],
capture_to_use['depth_image'], capture_to_use['boxes'],
ROTATION_ANGLES[capture_to_use['source']])
get_object_position_end = time.time()
print(f'system_cap_time: {capture_to_use["system_cap_time"]}, '
f'image_queued_time: {capture_to_use["image_queued_time"]}, '
f'processed_image_start_time: {capture_to_use["processed_image_start_time"]}, '
f'processed_image_end_time: {capture_to_use["processed_image_end_time"]}, '
f'get_object_position_start_time: {get_object_position_start}, '
f'get_object_position_end_time: {get_object_position_end}, ')

# get_object_position can fail if there is insufficient depth sensor information
if not world_tform_object:
continue

scores = capture_to_use['scores']
print(f'Position of object with confidence {scores}: {world_tform_object}')
print(f'Process latency: {time.time() - capture_to_use["system_cap_time"]}')
tag_cmd = get_go_to(world_tform_object, robot_state, params_set)
end_time = 15.0
if tag_cmd is not None:
if not options.test_mode:
print('executing command')
robot_command_client.robot_command(lease=None, command=tag_cmd,
end_time_secs=time.time() + end_time)
else:
print('Running in test mode, skipping command.')

# Shutdown lease keep-alive and return lease gracefully.
lease_keep.shutdown()
lease_client.return_lease(lease)
return True
except Exception as exc: # pylint: disable=broad-except
LOGGER.error('Spot Tensorflow Detector threw an exception: %s', exc)
# Shutdown lease keep-alive and return lease gracefully.
return False


if __name__ == '__main__':
if not main():
sys.exit(1)

Dans un scénario, les soldats chinois se tiennent de part et d'autre d'une porte tandis que le chien robot pénètre dans le bâtiment devant eux ; dans un autre, le robot tire une rafale de balles alors qu'il avance sur une cible. « Il peut servir de nouveau membre dans nos opérations de combat urbain, remplaçant nos membres pour mener la reconnaissance, identifier l'ennemi et frapper la cible pendant notre entraînement », rapporte la CCTV des propos d’un soldat chinois montrant le robot en train d'opérer.

L'année dernière, le Pentagone a expérimenté l'équipement de robots terrestres quadrupèdes avec sa carabine standard M4A1 de 5,56 x 45 mm, le fusil XM7 de 6,8 mm que l'armée américaine est en train d'adopter dans le cadre de son programme « Next Generation Squad Weapon », et même l'arme antichar légère M72 qui est en service au sein des troupes américaines depuis la guerre du Vietnam. Quelques semaines avant que CCTV ne publie ses images de chiens robots armés en action, le Marine Corps Special Operations Command (MARSOC) a révélé qu'il expérimentait l'ajout à ses propres chiens mécanisés de systèmes d'armes montés basés sur le système d'armes à distance SENTRY de l'entreprise de défense Onyx, basé sur l'intelligence artificielle.

Les responsables américains de la défense se sont empressés de souligner que le développement de robots-chiens armés était, à ce stade, purement expérimental, destiné à aider les planificateurs militaires à explorer le domaine du possible en ce qui concerne les applications potentielles de systèmes robotiques révolutionnaires dans un futur conflit, comme l'a déclaré un responsable de l'armée en août dernier. Mais comme les soldats de l'armée de terre effectuent des exercices d'assaut urbain avec des robots-chiens et que le corps des Marines envisage de plus en plus d'utiliser des quadrupèdes mécaniques pour renforcer les formations futures grâce à la robotique intelligente, l'armée américaine pourrait bien être obligée d'envisager sérieusement d'adopter des chiens robots armés pour le combat, avant la Chine.

Et vous ?

:fleche: Pour ou contre l'utilisation de l'IA à des fins militaires ? Si pour, à quel degré ou à quelles fins ? Si contre, pourquoi ?
:fleche: Êtes-vous surpris de la mise à contribution de l’intelligence artificielle sur les champs de bataille ?

Voir aussi :

:fleche: Des représentants de 150 entreprises de la tech plaident contre les robots tueurs autonomes, lors d'une conférence internationale sur l'IA
:fleche: USA : le sous-secrétaire à la défense soulève l'importance d'associer l'intelligence artificielle à l'armée, évoquant une future course aux armes
:fleche: Des experts en IA boycottent un projet de robots tueurs d'une université en Corée du Sud, en partenariat avec le fabricant d'armes Hanwha Systems
:fleche: Non, l'IA de Google ne doit pas être au service de la guerre ! Plus de 3000 employés de Google protestent contre la collaboration avec le Pentagone
:fleche: Robots militaires : les États-Unis testent l'efficacité de leurs prototypes dans le Pacifique, l'armée irakienne présente son robot mitrailleur