A digital COgnitive architecture to achieve Rapid Task programming and flEXibility in manufacturing robots through human demonstrations (DIGI-CORTEX)

数字认知架构，通过人体演示实现制造机器人的快速任务编程和灵活性（DIGI-CORTEX）

• 批准号: EP/W014688/1
• 负责人: John Oyekan
• 金额: $ 41.47万
• 依托单位: University of Sheffield
• 依托单位国家: 英国
• 项目类别: Research Grant
• 财政年份: 2022
• 资助国家: 英国
• 起止时间: 2022 至 无数据
• 项目状态: 已结题
• 来源: https://gtr.ukri.org/projects?ref=EP%2FW014688%2F1
• 关键词: digital; COgnitive; architecture; achieve; Rapid

The Made Smarter review identified that the UK is lagging behind in worker productivity and could benefit from the advent of new industrial digital tools (IDTs) such as novel intelligent technologies, connected devices, robotics and artificial Intelligence. It is estimated that IDTs could contribute an additional £630bn to the UK economy by 2035 and increase the manufacturing sector growth by between 1.5 and 3% per annum [1]. For example, there is a high demand for bespoke and personalised goods in high volume [2]. In order to meet this demand, manufacturing systems need to be highly flexible, adaptable and highly automated. Since most manufacturing SMEs make use of jobshops and contribute up to 15% of the UK economy [3], equipping them with robots that can learn a task rapidly and flexibly (similar to how a human can be rapidly trained to assemble new product lines) will enable SMEs to meet high order demands thereby improving UK PLC's export opportunities and UK's GDP. This proposal aims to investigate cognitive architectures that equips robots with the capability to rapidly learn new skills by passive observation of a human demonstrating a task to the robot and applying previously learnt skills to new task scenarios, thereby achieving task flexibility on the manufacturing floor. This opens up exciting possibilities. For one, it means that robots can be taught to do various tasks with no intensive programming required by a human. It also means that robots can be flexibly used to perform a wide variety of tasks thereby reducing the need for capital intensive, rigid and time-consuming manufacturing set ups.There is a gap in literature of applying digital mental models on robots for building in flexible and creative robots that can be flexibly and rapidly re-tasked for various tasks. Nevertheless, there is a growing realisation that creativity is needed in industrial robots of the future and that this could be achieved through providing them with mental models [4]. For the first time ever, this proposal investigates a cognitive architecture that embeds the human cognitive capabilities of mental simulation for creative problem solving on manufacturing robots and task structure mapping in a unified framework for the purposes of achieving rapid re-tasking (task flexibility) of industrial robots via passive human demonstrations. State of the art architectures (such as SOAR and ART-R) often make use of a prior task informed rigid procedural rules that make them less amenable for exploring rapid re-tasking on robots while techniques that use machine learning paradigms (e.g deep neural networks or reinforcement learning) that require lots of data and result in task specific applications. Furthermore, these techniques are yet to be successfully combined with the creation of digital mental models through envisioning and applied to varying tasks in manufacturing environments similar to those to be investigated in this proposal. In summary, the novelty of this proposal is in the application of robot envisioned digital mental models to support them in creativity and imagination of morphological informed solutions to problems encountered in manufacturing (and other sectors outside manufacturing) as well as to support the application of previously learnt skills to new similar tasks. This will lead to rapid re-tasking and task flexibility in robots. References:[1] J. Maier, "Made Smarter Review," 2017.[2] D. Brown, A. Swift, and E. Smart, "Data analytics and decision making," Inst. Ind. Res. Univ. Portsmouth, pp. 1-20, 2019, doi: 10.4324/9781315743011-9.[3] C. Rhodes, "Business Statistics," 2019.[4] J. B. Hamrick, "Analogues of mental simulation and imagination in deep learning," Current Opinion Behavioral Science, vol. 29, pp. 8-16, 2019.

Made Smarter评估发现，英国在工人生产力方面落后，可以从新的工业数字化工具（IDT）的出现中受益，如新型智能技术，连接设备，机器人和人工智能。据估计，到2035年，IDT可以为英国经济贡献额外的6300亿英镑，并使制造业每年增长1.5%至3%[1]。例如，对大量定制和个性化商品的需求很高[2]。为了满足这一需求，制造系统需要高度灵活、适应性强和高度自动化。由于大多数制造业中小企业利用车间，并贡献了英国经济的15%[3]，为他们配备可以快速灵活地学习任务的机器人（类似于人类如何快速培训组装新产品线）将使中小企业能够满足高订单需求，从而改善英国PLC的出口机会和英国的GDP。该提案旨在研究认知架构，使机器人能够通过被动观察人类向机器人演示任务并将以前学到的技能应用于新的任务场景来快速学习新技能，从而实现制造车间的任务灵活性。这开辟了令人兴奋的可能性。首先，这意味着机器人可以被教导做各种任务，而不需要人类进行密集的编程。这也意味着机器人可以灵活地用于执行各种各样的任务，从而减少对资本密集型，刚性和耗时的制造设置的需要。在将数字心智模型应用于机器人以构建灵活和创造性的机器人的文献中存在空白，这些机器人可以灵活地快速重新分配各种任务。然而，人们越来越意识到，未来的工业机器人需要创造力，而这可以通过为它们提供心智模型来实现[4]。这是有史以来第一次，该提案研究了一种认知体系结构，该体系结构将人类的心理模拟认知能力嵌入到制造机器人和任务结构映射的统一框架中，以通过被动的人类演示实现工业机器人的快速重新分配任务（任务灵活性）。最先进的架构（如SOAR和ART-R）通常利用先前的任务通知严格的程序规则，这使得它们不太适合探索机器人上的快速重新任务，而使用机器学习范例（例如深度神经网络或强化学习）的技术需要大量数据并导致任务特定的应用。此外，这些技术还没有成功地结合数字心理模型的创建，通过设想和应用于不同的任务，在制造环境中类似于那些在本提案中进行调查。总之，该提案的新奇在于机器人设想的数字心智模型的应用，以支持它们在形态学信息解决方案的创造力和想象力中对制造业（和制造业以外的其他部门）遇到的问题的解决方案，以及支持将以前学到的技能应用于新的类似任务。这将导致机器人快速重新分配任务和任务灵活性。参考文献：[1] J. Maier，“Made Smarter Review”，2017年。[2]D.布朗，A. Swift和E. Smart，“Data analytics and decision making，”Inst. Ind. Res. Univ.朴茨茅斯，pp. 1-20，2019，doi：10.4324/9781315743011-9. [3]C.罗兹，“商业统计”，2019年。[4]J. B. Hamrick，“深度学习中的心理模拟和想象力的类似物”，Current Opinion Behavioral Science，第29卷，第29页。2019年8月16日。

项目成果

Graph-based semantic planning for adaptive human-robot-collaboration in assemble-to-order scenarios

• DOI: 10.1109/ro-man57019.2023.10309425
• 发表时间: 2023-08
• 期刊: 2023 32nd IEEE International Conference on Robot and Human Interactive Communication (RO-MAN)
• 作者: Ruidong Ma;Jingyu Chen;John Oyekan

An Improved Hybrid Multi-Objective Particle Swarm Optimization to Enhance Convergence and Diversity

• DOI: 10.1145/3583133.3596365
• 发表时间: 2023-07
• 期刊: Proceedings of the Companion Conference on Genetic and Evolutionary Computation
• 作者: Nazrul Islam;J. Oyekan

A learning from demonstration framework for adaptive task and motion planning in varying package-to-order scenarios

• DOI: 10.1016/j.rcim.2023.102539
• 发表时间: 2023
• 期刊: Robotics Comput. Integr. Manuf.
• 作者: Ruidong Ma;Jingyu Chen;J. Oyekan

A deep multi-agent reinforcement learning framework for autonomous aerial navigation to grasping points on loads

• DOI: 10.1016/j.robot.2023.104489
• 发表时间: 2023-07
• 期刊: Robotics Auton. Syst.
• 作者: Jingyu Chen;Ruidong Ma;J. Oyekan