Repository of DUTH

Data-driven neuromorphic decision-making for cognitive robotic navigation (Doctoral thesis)

Κούτρας, Δημήτριος/ Koutras, Dimitrios

The objective of this thesis is to develop algorithms and methodologies that enable robotic decision-making processes to closely emulate human behavior. Specifically, it addresses the challenges of robotic navigation and decision-making by utilizing neuromorphic architectures, such as neural networks, trained on heterogeneous datasets. This approach lays the groundwork for creating cognitive machines that can interact intelligently with real-world environments, moving toward the realization of General Artificial Intelligence (GAI). The thesis is organized into two main pillars. Pillar A establishes the theoretical foundations necessary for achieving Generalized Robotic Intelligence (GRI). It introduces the mathematical and algorithmic structures upon which this research is built. Chapter 3 presents an algorithm leveraging advances in machine vision to estimate object positions, enhancing the situational awareness of a swarm of Unmanned Aerial Vehicles (UAVs). By implementing a real-time learning strategy based on the well established Cognitive Adaptive Optimization (CAO) family, this algorithm significantly improves UAV coordination. However, limitations were identified, particularly in CAO’s need for reinitialization in new environments. To address this, reinforcement learning (RL) was integrated, leading to the development of Mars-Explorer, an environment detailed in Chapter 4. Mars-Explorer serves as a benchmark within the openai-gym suite, reformulating exploration and coverage tasks in an RL-compatible framework. This framework demonstrates how RL-based neural networks function analogously to computer RAM, storing and adapting knowledge. Testing RL algorithms in Mars-Explorer revealed critical challenges, prompting the creation of ACRE in Chapter 5. ACRE addresses issues in integrating additional reward signals, enabling more consistent access to environmental rewards through a replay buffer. Pillar B focuses on the practical application of the theoretical frameworks in real-world scenarios. Robotics inherently demands real-life implementation, necessitating a custom robotic platform to fully realize the thesis’s objectives. Chapter 6 details the design and construction of a custom Unmanned Ground Vehicle (UGV), tailored for flexibility and equipped to support state-of-the-art algorithms. The companion computer onboard the UGV provides unparalleled flexibility, allowing the execution of complex software configurations beyond those available in off-the-shelf solutions. Chapter 7 culminates the thesis with CRN: Cognitive Robotic Navigation, a neuromorphic framework that enables autonomous robots to navigate complex environments using only visual input. Leveraging advanced visual transformers and generative AI, CRN interprets RGB images to produce a robust decision-making framework capable of generating and assessing navigational paths. Inspired by the cognitive mechanisms observed in humans and animals, CRN adopts naturalistic navigation strategies, utilizing perceptual cues to interpret and interact with the surrounding environment effectively.