Abstract
Modern manufacturing relies substantially on robotic systems. Robots can not only help reduce the workload on humans, but they can also improve precision, accuracy, reduce downtime, and help improve the overall efficiency of manufacturing setups. Industry 5.0, which focuses on a human-centric approach and Flexible Smart Manufacturing Systems (FSMS), needs smart robotic systems that can easily handle changes, implying more generalizability. Sensorimotor planning in robots faces significant challenges mostly owing to the variability in scenarios, settings, and tasks at hand. Hence, improving generalizability and reducing the dependence on human oversight is a key need for robotic systems of the future.
This thesis focuses on novel methods aiming to reduce the load on human operators by developing robotic systems that can handle changes in environment, tasks, and systems—or a combination thereof. First, this research presents a hierarchical approach to combine Learning from Demonstration (LfD) with Reinforcement Learning (RL) to handle robotic assembly problems, devising an approach that can handle variations in tasks. The research further builds upon this to incorporate a novel Dynamic Movement Primitive (DMP)-based method to allow robotic manipulators to handle variations in environments. The ability of the proposed research to handle static and dynamic obstacles is leveraged for multi-manipulator motion planning. Coordination is further enhanced by leveraging a Multi-Agent Reinforcement Learning (MARL) technique. Lastly, the research focuses on employing generative learning-based methods to handle multi-agent motion planning. Specifically, the proposed research leverages diffusion models to generalize the motion planning for multi-agent systems where the number of agents is changing dynamically.
Through this systematic progression, the research advances the state-of-the-art in generalizable robotic systems, contributing to the realization of autonomous manufacturing systems capable of operating effectively in dynamic, unstructured environments characteristic of next-generation industrial applications.
