現今大部份的群體機器人皆不具有良好的感知能力，導致了群體機器人不易達成協同控制，而無法有效地完成工作目標。目前市面上已有不少具有基礎影像處理能力且低成本的單板電腦。因此，本計畫將利用單板電腦實現一具有影像視覺及群體學習的群體機器人。首先，為了確保有足夠的影像運算能力及穩定的控制，我們將開發一具二核心的低成本個體機器人，群體中每個個體皆可進行影像運算，也可進行機器人的動作控制。為了確保群體能進行訊息的分享與學習，我們將於一個（或多個）個體中啟動無線網路服務，以確保無線通訊可以完全覆蓋全部的個體。個體間的訊息傳遞將利用機器人運算系統（ROS）所提供的發佈/訂閱的機制來進行個體間的訊息溝通，同時，利用黑板結構來達到個體之間的資料、及訊息的分享。在群體機器人協調、合作控制方面，基於Vicsek模型，我們將使用分佈式協同控制，並以Lyapunov定理證明群體動作的穩定性，同時，於群體機器人合作學習方面，我們將提出一增強式學習方法，藉由全部個體的經驗分享、獎勵或處罰的回饋來修正群體的動作。為了確保群體與環境中的位置，每個個體將使用SIFT演算法進行特徵比對，以找到在它視覺範圍內的其它個體的位置與相對距離，及於部份個體中將利用Gmapping定位演算法進行群體的定位。最後，我們將以群體機器人的防撞、聚集、及覆蓋任務來進行模擬及實驗驗證。
Nowadays, the cognitive ability for most of swarm robots is not good enough for achieving an assignment by using collective control. Lots of cheap single-board computers which are capable of image processing hit market in recent years. Therefore, this project aims at using a single-board computer to design a swarm robot capable of not only vision computation but also modifying their actions from all of individual robots. The developed low-cost individual robot has two cores, which one is applied for image processing and the other one is applied for robot control . Any individual of the swarm robot has ability of independently finishing actions according to its own vision sensor and supersonic sensors. In order to ensure that wireless communication is always connectable for all of the individuals, we will choose one (or more) individual which enables wireless internet service as a mobile wireless access point to cover the area where the swarm robot located on. The individuals are able to communicate with one another according to their published or subscribed topics, and, furthermore, a blackboard structure is able to record the actions and information from the individuals and provides a proper action response for them. This project will adopt Vicsek model to represent the dynamic of the swarm robot. According to the Vicsek model, the stability of connective control for the swarm robot can be proved by the Lyapunov theory. In addition, we use reinforcement learning for the swarm robot, where each of the individuals shares their information and experience to the swarm and the swarm returns award or punish back to the individuals to modify actions. Depended on SIFT image operation, each of the individuals finds distances and direction between others in front of the individual, and a part of the individuals are assigned to execute Gmapping localization algorithm in order to correctly locate the swam. Finally, we will conduct the designed swarm robot in three cases: a collision-free assignment, a gather assignment, and a coverage assignment.
