No abstract

This paper describes the architecture and implementation of a heterogeneous team comprising unmanned ground vehicles and blimp robots capable of navigating unknown subter- ranean environments for search and rescue missions. The ground vehicles are equipped with a range of sensors for accurate perception, localization, and mapping. The blimps feature a long flight duration and collision tolerance when traversing uneven terrain. The design of the system was meant to satisfy the requirements of the Defense Advanced Research Projects Agency (DARPA) Subterranean Challenge in terms of perception capability and autonomy. To facilitate navigation through smoke-filled spaces, we employed novel millimeter wave radar to enable cross-modal representations for integration via deep reinforcement learning. The autonomy of the proposed scheme was augmented using simulations to train deep neural networks, thereby allowing the system to perform sequential decision-making for collision avoidance and navigation toward a specific goal. The navigation system was evaluated in the DARPA SubT Urban Circuit, and quantitative localization results and recovery strategy from failures was discussed. The proposed communication system comprises mesh WiFi with XBee (ZigBee network with XBee radios) and ultra-wideband (UWB) communication modules as well as spherical nodes that can be shot out like a cannonball and miniature cars deployed as mobile nodes. The propagation and radio signal strength index of various modules were evaluated using data collected during field tests in order to overcome the uncertainties of subterranean environments, including non-line-of-sight propagation, multipath propagation, and fading reception. We also discuss the lessons learned during this project and reflect on future plans.

Colab CPU, a Colab GPU (which is either a Tesla P100 or 230 a Tesla T4), an Nvidia Jetson Xavier and a Nvidia Jetson 231 Nano: the server GPU is the fastest followed by the Xavier, 232 the Nano and the server CPU. The same ranking occurs for 233 the scaling ratio: D3(1-5) is 3.2 times slower than D0(1-234 5) with the server GPU, 5.3 times slower with the Xavier, 235 6.3 times slower with the Nano and 7.9 times slower with 236 the server CPU. Therefore, a more powerful GPU has a 237 double benefit: smaller runtimes and smaller scaling ratios; 238 however, the full-fledged GPUs used in large clusters cannot 239 be installed on underwater robots due to energy, space and 240 weight constraints. 241 B. On Marine Debris Datasets 242 Subsequently, we trained our improved EfficientDets on 243 the first 1200 samples of the Trash-ICRA19 [11], which is 244 a publicly available annotated dataset for object detection 245 targeting three classes: "bio" for biological items, "plastic" 246 for plastic items and "rov" for remotely operated vehicles. 247 The results are in Table VI, where 35 epochs have been 248 chosen for demonstration purposes; when training a model 249 for deployment on an AUV, the number of epochs should 250 be increased in order to reach the learning saturation. With 251 35 epochs, we reached 47.4% AP with D0 and 49.6% AP 252 289 4; most of the samples of WPBB are taken at more than 290 a meter of distance from the object, hence most of the 291 pixels are background; some samples of WPBB have low 292 resolution because captured while the drone was steering. 293 All these factors make the WPBB dataset both challenging 294 and realistic. Examples of detections are shown in Fig. 4 295 with different values of confidence threshold τ , which is an 296 hyperparameter meaning that detections with a confidence 297 lower than τ are ignored: the plastic bag is detected by all 298 the four detectors with high confidence, but its detection is 299 relatively easy since it is well visible; in contrast, one of the 300 two floating bottles at the top is detected only by D0 with 301 just a 43% confidence. 302 C. Evaluations on Simulator 303 Given that underwater data were expensive to collect, 304 we further carried out reproducible simulation evaluations 305 using the proposed D2(1-5). The simulation in Unity 3D 306 engine provided realistic visual effects and labor-free ground 307 truth labeling. The general setup is similar to our previ-308 ous work [27]. The system included two simulators: Unity 309 (virtual reality (VR) side) and Gazebo (robot side). In the 310 current study, the VR side was involved to transfer real-311 time realistic images rendered by the Unity 3D engine to 312 the object detector. Both Unity and Gazebo included its own 313 virtual vehicle, i.e., a boat equipped with an underwater 314 camera. The robot side was designed to switch either via 315 the Gazebo simulation or a real robot. The robot side 316 included an object detection module running the D2(1-5) 317 model, a navigation and control module pe...

No abstract