計畫名稱：以模糊演化式卡曼過濾器設計具適應性與重新規劃能力之自主性機器行動機制
執行起迄：2011/08/01~2012/10/31
總核定金額：499,000元
中文摘要：人工智慧(AI)與機器人的關係密不可分，要讓AI融入機器人，使之能自主地執行任務，向來是機器人研究的重要目標之一。本計畫的動機源於如何提升機器人在長期運轉、無人類可以協助的情形下，如何以低資源消耗、適度自主性、環境適應性、可重規劃機體等功能，進行自主行動與故障排除的能力。本計畫從AI的角度出發，檢視近代AI技術在智慧型機器人的執行模式，並重新設計如何在低成本、低耗能、低資源的硬體環境下，機器人仍保有適當的智慧與自主性。另外，機器人如何以AI的方式驅動可重規劃硬體(re-programmable hardware)進行自主容錯的任務要求，也是計畫重點。本計畫以三年為規劃：第一年計畫著重在「環境與感測」，研究如何從低功率、低成本簡易型感測資料，期待建立「即使感測器不精確」機器人仍能有智慧地建立環境感測模型。計畫以模糊邏輯(fuzzy logic)理論建立感測器的測量模型，將多個感測器的模糊訊號交叉比對作為訊號融合(sensor fusion)的修正依據，以演化式卡曼過濾模式(evolvable Kalman filtering)的方式，搭配模糊比對(fuzzy pattern matching)技術建立環場感測模型。第二年計畫著重在「環境與驅動」，以行動規劃為主要任務，期待建立「即使環境是未知的且制動器結果不良」機器人仍能有智慧地執行已設定的任務。計畫採用有限個環境感測模型(第一年成果)與行動對應的模糊關係，並使用演化式增強學習法(evolutionary reinforcement learning)修正行動。第三年將前二年成果整合，著重在「故障與容錯」，期待建立「即使機器發生故障或損毀，機器人仍能有智慧地進行硬體與任務計畫的重新規劃，使任務不中斷」。計畫以遞增式增強學習與遺傳程式規劃技術建立具故障容錯能力的演化式硬體機器人任務行動規劃。本計畫考慮以小型機器人為實驗平台，所有AI演算機制均以可重規劃式硬體(FPGA)實現。故需考慮計算資源的有限性、計算過程的收斂性、計算結果的精確度。
英文摘要：Artificial intelligence (AI) is an important topic in robotics research. AI is one of the core techniques of autonomous, or smarter, robots. This project motivates from the needs of robots equipped with low-level and low-accuracy sensors and actuators and limited computational resources can still operate autonomously. Additionally, such robots need to have an ability of fault-tolerance. In this project, we will investigate the recent AI techniques that can be applied to autonomous robots and study how these techniques can be employed in low-power consumption autonomous robots with low-level sensors/actuators and limited computational resources. Regarding the fault-tolerance ability, how these AI techniques can be implemented in a resource-constrained reconfigurable hardware will be an issue. Here we conduct a three-year project for these issues. In the first year of the project, we study how accurate environmental models can be re-constructed from a set of low-level, low-cost, inaccurate sensor. We use evolutionary Kalman filtering with fast fuzzy pattern matching for multi-senor fusion. The second stage of the project focuses on how robots drive low-level, low-cost, inaccurate actuators autonomously in an unknown environment. The techniques to be used include fast fuzzy genetic programming and evolutionary reinforcement learning. In the third stage of the project, fault-tolerance and non-stop mission execution are the main theme. Sensors and actuators may be out of order during the mission execution of robots. The ability of reconfiguration on hardware connections and software parameters is essential in such a scenario. For this purpose, we investigate how the incrementally evolutionary genetic programming technique can be implemented in a reconfigurable hardware. All these techniques are implemented not only in software environment but also in an FPGA-based reconfigurable hardware so that the running time to convergence, resource-consumption, the accuracy of resolutions of these AI techniques must be re-investigated. This project is one of few projects that integrate AI techniques of both hardware and software for robotics research.