Human-inspired dexterous manipulation of deformable objects: towards economically sustainable robotic textile recycling

Abstract

Textile manufacturing and disposal are among the most environmentally damaging product life-cycles in the early 21st century. For example, clothing disposal in landfills can cause significant damage and is expanding due to fast-fashion business practices. In addition, as clothing decomposes it emits the greenhouse gas methane, and toxic chemicals, including dye, can leak into the soil. Recycling discarded clothing by chemically breaking it down to produce new materials, such as lyocell, has been suggested to combat some of these issues. However, it is rare to see these technologies applied on a large scale, partly due to the extensive labour required to sort discarded clothing to the appropriate recycling process. This thesis examines the utilisation of robots to mitigate labour demands, focusing on pick-and-place applications involving deformable object manipulation. Robotic manipulation of textiles remains challenging as fabric displays a unique physical behaviour due to clothing’s anisotropic nature and non-linear mechanical response. Furthermore, garments can exhibit various colours, shapes, forms and textures, making physical manipulation and visual interpretation a series of complex and multifaceted tasks. This thesis addresses dexterous manipulation, recognising that clothing can present itself in states requiring skilful manipulation, and handling garments can require intrinsic dexterous skills. As a result, several academic projects have developed end-effectors to manipulate fabric. However, these solutions are not generalised and target specific manipulation pipelines such as folding a shirt or grasping flattened material. While previous research uses heuristic human observations to inspire robotic gripper designs, these approaches do not extensively explore human behaviour and morphology in a robotic context. This observation is what inspires the research described in this thesis. First, an investigation using anthropomorphic taxonomies defines the range of skills necessary for a generalised fabric pick-and place solution. This device is then modelled using classical mechanics and fabricated. Then, data driven approaches in reinforcement learning are applied to build behaviours that enable the end-effector to robustly execute environmentally constrained behaviour. Finally, this thesis provides a discussion of the conducted research and an overview of future applications.

Date of Award	2024
Original language	English
Supervisor	Damith HERATH (Supervisor) & Roland Goecke (Supervisor)