A Machine Learning Technique for Motion Planning in Articulated Robots

Abstract

The integration of human-like motion into robotic systems has emerged as a pivotal research frontier, driven by the aspiration to create robots capable of nuanced interactions in real-world scenarios. This thesis delves into the development and implementation of a novel solution, leveraging state-of-the-art motion planning algorithms and neural network based technologies to instill human-like qualities in robotic motion. The research unfolds against the backdrop of contemporary challenges in the field of robotics, particularly the need for motion planning algorithms (MPAs) that not only navigate collision-free paths but also emulate human-like movements. In the landscape of motion planning, traditional algorithms like rapidly exploring random trees (RRT) and its variants have proven effective, yet their computational complexity becomes a bottleneck in higher-dimensional problem spaces. To address this limitation, the motion planning networks (MPNet) paradigm is introduced, utilizing a neural network approach with point cloud representations to navigate the intricacies of higher-dimensional environments. The computational efficiency of MPNet is harnessed to overcome challenges associated with dimensionality, offering a promising avenue for generating collision-free paths. The central problem addressed by this research is the lack of human-likeness in the paths generated by MPAs, limiting their applicability in tasks that demand human-like motion. Prior attempts to imbue human-likeness often relied on datasets recorded from human movements, leading to unreliable and constrained solutions. In response, the proposed solution adopts a hybrid approach, combining the strengths of MPNet and Artificial x VFRRT. The latter is chosen for its ability to generate human-like paths, albeit with limitations in higher-dimensional problems. The research methodology unfolds in distinct phases, beginning with the development of a reliable human-like motion dataset. The Data Generator (DG) module orchestrates this process, employing an Artificial VFRRT-based motion planner within the Kautham simulation tool. The dataset, characterized by dynamic path generation strategies and environmental diversity enhancements, forms the foundational building block for subsequent modules. The Data Encoder (DE) module steps in to transform raw obstacle representations into a point cloud format compatible with MPNet training. This adaptive encoding ensures usability and efficiency, setting the stage for the Human-like MPNet (HLMPNet) module. HLMPNet marks a paradigm shift in motion planning architectures, dynamically adapting learning parameters through an iterative process informed by human-likeness evaluations. This module is designed not only to replicate human motions but to refine and adapt its behavior based on nuanced feedback. The Human-Likeness Evaluator module acts as the discerning judge in the evaluation framework, quantifying the authenticity of generated paths. Its role in continuous learning and optimization ensures that HLMPNet evolves towards increasingly authentic and nuanced human-like motion planning. The significance of this research lies in its practical applications across various domains, including real-world human-robot collaboration, user-friendly interfaces, efficient and safe robotic operations, enhanced experiences in entertainment and services, and improved assistive and rehabilitation technologies. The proposed solution offers a holistic approach to addressing the challenges of human-like motion in robotics, contributing to the ongoing evolution of robotic systems in diverse applications