Hybrid Robotics for Future Reconfigurable Manufacturing

目前制造业中的机器人系统主要基于“零件到过程”的概念，其中零件被交付并精确定位在固定的机械手机器人单元中。大多数这样的系统缺乏灵活性，执行严格编排的任务，不适合可重构制造。例如，汽车行业的焊接和喷漆应用。自动化工业的最新发展导致了协作式移动机械臂机器人的出现，特别是将机械臂的能力与移动平台相结合。这些平台为制造业实现“过程到部件”的理念提供了独特的机会。例如，该过程现在可以在制造点被带到零件上，从而实现未来的可重构制造系统。本研究旨在开发基础技术，使移动机器人操纵者的实时同步操作和自主推理能力在可重构设置中充分发挥潜力。在这里，它寻求使用深度强化学习技术来解决逆轨迹规划和推理问题，结合先进的传感器技术，如3D视觉，末端执行器力-扭矩控制，激光雷达等，以最少的人力投入生成机器人程序。例如，一旦生产线重新配置，机器人将进行自主推理，以了解要执行的新任务集，并确定如何以及何时在没有人为干预的情况下执行这些任务。然而，在现实世界中成功训练强化学习算法可能需要多年的训练数据、尝试和研究人员的时间，并且存在重大的事故风险和对机器人硬件及其环境的破坏。因此，本研究建议使用“模拟到真实”的迁移学习方案来训练给定的强化学习算法。在模拟到真实中，算法的训练阶段是在虚拟环境中进行的，通过在物理模拟器中对机器人及其操作域进行建模。然后，训练好的算法将使用域适应或类似的技术转移到现实世界的机器人上，并进行实时操作测试。这种方法减少了在现实世界中训练机器人的时间、精力、成本和风险。然而，模拟到真实技术可能面临与仿真环境中的物理建模和多领域之间准确的知识转移相关的挑战

Current robotic systems in manufacturing are predominantly based on the "Part-to-Process" notion, where the part is delivered and precisely positioned within a fixed manipulator robotic cell. Most such systems lack flexibility, execute a strictly choreographed task, and are unsuitable for reconfigurable manufacturing. E.g., welding and spray painting applications in the automotive sector. The recent developments in the automation industry led to the emergence of collaborative mobile manipulator robots, particularly combining the capabilities of robotic arms with mobile platforms. These platforms offer a unique opportunity to realise the idea of "Process-to-Part" in the manufacturing industry. For example, the process can now be brought to the part at the point of manufacture, enabling future reconfigurable manufacturing systems.This research aims to develop fundamental technology to enable mobile robotic manipulators' real-time synchronised operation and autonomous reasoning capabilities to realise the full potential in reconfigurable settings. Here, it is sought to use Deep Reinforcement Learning techniques to solve the inverse trajectory planning and reasoning problem combined with advanced sensor technologies such as 3D vision, end effector force-torque control, LiDAR, etc., to generate robotic programs with minimal human input. E.g., once the production line is reconfigured, the robot will do autonomous reasoning to understand the new set of tasks to be executed and establish how and when to perform them without human intervention.However, successful training of a reinforcement learning algorithm in the real world may require years of training data, attempts, and researcher time and poses a significant risk of accident and damage to the robotic hardware and its environment. Therefore, this research suggests using the "Sim-to-Real" transfer learning scheme to train a given reinforcement learning algorithm. In Sim-to-Real, the algorithm's training phase is performed in a virtual environment by modelling the robot and its operating domain within a physics simulator. The trained algorithm will then be transferred to the real-world robot using domain adaptation or a similar technique and tested for real-time operation. This method reduces the time, effort, costs and risks of training the robot in the real world. However, the Sim-to-Real technique may have challenges associated with physics modelling in the simulation environment and accurate knowledge transfer between multiple domains