دانلود پایان نامه شبکه های عصبی عمیق پویای تطبیقی برای سیستم های خودمختار

Autonomous operation of robots and sensors in remote, unknown environments require a rigorous interdisciplinary approach bridging computer science, mathematics, and engineering. A central challenge in deploying full autonomy stacks is managing the cyclic flow of data from raw acquisition to intermediate perception and final predictions. While deep neural networks (DNN) enable advanced perception and logic, they are computationally demanding and static in nature. This clashes with the resource demands of remote systems which are constrained by embedded processors, lightweight sensors, and limited energy reservoirs. Dynamic DNNs (DDNN) provide great promise in mitigating this bottleneck by enabling runtime branching and scaling, but they often lack context awareness. Presented are end-to-end methodologies and applications focused on the development of adaptive DDNNs (ADDNN), which are demonstrated in novel frameworks to augment perception and navigation pipelines with the ability to intelligently select when and how to branch and scale. Ultimately, this research demonstrates that ADDNNs can break the traditional trade-off between efficiency and accuracy, by simultaneously, and significantly, improving both. Evidence to support this claim is collected by first developing a robust hardware-in-the-loop (HIL) testbed, that combines high-fidelity simulation (Microsoft AirSim) with deployment on edge computing hardware (NVIDIA Jetson). This setup rigorously benchmarks inference latency, power consumption, and energy expenditure, against navigation accuracy. A flexible navigation framework is designed and deployed to this testbed. At its core, is a generalized mid-level policy that uses a combination of rule-based, initiation learning, and deep reinforcement learning (DRL) algorithms to generate generalized motion actions in response to onboard sensor observations. A key strength of the mid-level policy is in its ability to be deployed to various real-time controllers, closing the loop between sensing and actuation for varying robot embodiments. The modular structure of the hierarchical navigation framework allows for varying sensor configurations and environmental complexities with little retraining and no fine tuning. This is further augmented with a high-level policy by integrating cutting-edge multimodal large language models (MLLMs), which are used to supervise and improve the performance of the entire framework in a zero-shot manner. Finally, considerations are taken into artifacts that are inherent with real-world scientific data that is processed through a combination of onboard and offboard computing, for two downstream tasks: modeling evapotranspiration in watershed environments, and autonomously classifying geological samples for Martian and lunar missions. Such considerations are taken in poor signal-to-noise ratios, missing data due to instrumental failure and human error, uncertainty quantification and reduction, and heterogeneous sensor fusion.