Empowering vision-guided robotic manipulation tasks with machine learning - with example applications to electric vehicle battery dismantling

Aflakian, Ali ORCID: 0000-0001-9777-7067 (2024). Empowering vision-guided robotic manipulation tasks with machine learning - with example applications to electric vehicle battery dismantling. University of Birmingham. Ph.D.

Aflakian2024PhD_Redacted.pdf
Text - Redacted Version
Available under License All rights reserved.
Download (34MB)

Abstract

This thesis is focused on the integration of advanced Artificial Intelligence (AI) techniques with computer vision methods to enhance the capabilities of robotic manipulation systems. The research application focuses on the disassembly of Electric Vehicle (EV) battery packs using robotic systems. As the demand for EVs continues to rise, the need for sustainable, and autonomous battery recycling grows, due to the limited lifespan of these batteries. Employing robots in this process is crucial as it guarantees not only cost-effectiveness and precision but also mitigates hazardous exposure for human workers. However, the disassembly of EV batteries is a challenging task for the robots. Firstly because of the diversity in battery types, shapes, and sizes and the absence of standardized formats across manufacturers. Secondly, robotics and automation in manufacturing typically operate within structured environments, doing repetitive tasks on known objects in fixed positions. However, adapting robots to handle different objects and unpredictable situations is challenging, necessitating innovative AI-based robotic solutions that can adapt to the diverse nature of battery packs while ensuring efficient and safe disassembly operations. To achieve these objectives, this study aims to develop control strategies that enable robots to perform precise and efficient manipulation tasks in dynamic and complex environments. The thesis is organized into several chapters, each addressing specific aspects of this comprehensive goal. The thesis begins by explaining the problem statement and difficulties related to the safe recycling of Lithium-ion batteries. It is certainly the case that challenges with disassembly procedures require automated solutions, with robots being a key component. Thereafter, a general introduction to the field, highlighting the need to enhance adaptability, precision, and efficiency in robotic systems through the integration of AI and vision-integrated techniques is presented. The objectives of the research are also provided, preparing the setting for the following chapters. The second chapter focuses on optimizing hybrid Visual Servoing (VS) approaches. Conventional methods in 2D, 3D, and hybrid VS are known to have limitations such as convergence issues, sub-optimal trajectories, and singularities. We carefully studied the behaviour of each approach in controlling specific components and proposed a decoupled hybrid VS approach. This approach integrates the functionalities of both 2D and 3D VS methods, minimizing their shortcomings while leveraging their distinct strengths. Adaptive gains, task sequencing, damped least squares, and projection operator techniques are integrated into the proposed method, each addressing a specific challenge to enhance the overall efficacy of the VS. Since the proposed method introduced computational complexities, we used neuro-fuzzy neural networks to predict and model the behaviour of the proposed VS approach. Then the proposed method is compared with traditional approaches, demonstrating improved efficiency and robustness. In the third chapter, the project investigates the integration of Reinforcement Learning (RL) techniques with VS. This involves expanding the AI intervention beyond image spaces to joint spaces of the robot. The main contribution of this chapter explores the incorporation of deep RL algorithms with the data of several controllers during training to enhance the performance of the trained policy. The proposed approach improves training by dynamically constraining the agent’s action spaces using several controller demonstrations, allowing for the learning of robust policies adaptable to various scenarios, and reducing the risk of sub-optimal solutions during training by utilizing the knowledge of mathematically proven control methods. Additionally, the proposed strategy can be combined with other techniques to enhance the training process. This approach results in a case study demonstrating a 51% reduction in training time, improved convergence rates, and enhanced controllability of the trained agent. Thereafter, the performance of the controller was demonstrated through comparisons with traditional VS methods. The fourth chapter extends the project’s scope to contact-rich manipulation tasks. A framework was introduced that employs RL to address the challenges posed by uncertain and complex environments. By utilizing a curriculum-based domain randomization approach, the robotic system’s robustness to uncertainties is improved, enabling the successful execution of compliant path-following in a range of conditions. To accelerate the RL training and reduce the problem of getting stuck in sub-optimal solutions, we proposed a strategy using human demonstrations. The data from human demonstrations in completing the desired task was gathered across various surfaces with various friction and stiffness. Thereafter, this data is used to create a 3D shape that includes all the demonstrated trajectories. This shape helps to limit where the RL algorithm should search for solutions. Notably, this strategy differs from imitation learning in that the agent is not required to imitate any particular behaviour. Therefore, it can accommodate even imperfect demonstrations, as the RL policy will correct the agent’s behaviour during the training. Chapter five underscores the practicality and adaptability of the developed techniques across a range of scenarios. This chapter presents a comprehensive exploration of various experiments and supplementary works carried out during the PhD project. These include tasks such as sorting, utilizing different grippers and robots, teleoperation, employing haptic devices, and applying computer vision methods such as deep learning for object detection, a model-based tracker for tracking predefined models, and the proposed VS approaches. The final chapter, the conclusion, summarizes the key findings, contributions, limitations, and implications of the research. It reflects on how the integration of AI and VS has led to advancements in robotic manipulation capabilities. This chapter also discusses potential future directions for research and emphasizes the broader impact of the project’s outcomes on robotics and automation. Overall, this PhD aims to bridge the gap between advanced computer vision, AI, and traditional robotic manipulation techniques, ultimately contributing to more efficient, adaptable, and robust manipulation systems capable of performing in complex real-world environments.

Type of Work:

Thesis (Doctorates > Ph.D.)

Award Type:

Doctorates > Ph.D.

Supervisor(s):