Introduction With the rapid development of microelectromechanical systems (MEMS), low-power embedded systems, and communication technologies, miniature sensors with sensing, computing, and wireless communication capabilities have been widely applied. These sensor networks, composed of wireless miniature sensors, can collaboratively monitor, sense, and collect information about various environments or objects within the network's distribution area in real time, process this information, and transmit it to users who need it. This is what BusinessWeek recognized as one of the 21 most influential technologies of the 21st century—Wireless Sensor Networks (WSN). In a massive 20-story hospital with thousands of rooms, how can an elderly or disabled patient, just entering the building and sitting in a smart car, easily reach their desired room? We are attempting to provide an "indoor automatic navigation system" that requires no external intervention—called the "Wireless Mosaic Eyes (WiME) system." In short, it is a wireless sensor network inspired by biological behavior, providing behavioral control for smart cars through a large number of wireless nodes distributed in the air; therefore, it is a robot navigation system implemented using a wireless sensor network. WiME involves two routing problems: one is robot path planning in geospatial space, and the other is information communication routing between distributed communication nodes. Compound eyes can act as electronic beacons in the robot navigation process; the wireless compound eye network can be considered a topological map describing geospatial space, geographic path planning, or it can be simplified to path planning in a network topology map. Therefore, in WiME, spatial path planning and information communication routing can work in exactly the same way, and path planning will be based on the semantic definitions of each distributed node. To achieve robot navigation with a complete map on nodes in a system like WiME using wireless sensor network technology, this design uses a single-step direction query path storage and query system. To further reduce the amount of path information data in resource-constrained wireless sensor nodes, a Bloom filter is used to compress and store each group in the WiME design. Additionally, since path information may be dynamically created, each sub-table is designed as a counting Bloom filter to meet the requirements of frequent modifications. 2. WiME System 2.1 WiME's Bio-Inspired Autonomous Robots are typically equipped with sensors that can perceive the environment and move autonomously. These sensors can be considered the robot's organs, and the information they perceive is centrally processed by an onboard computer. This centralized architecture faces three major challenges: sensor processing is not robust in real-world environments; computationally intensive processing algorithms operate in dynamic environments; and understanding scenes in unstructured environments is extremely difficult. However, compared to vertebrate eyes, the eyes of even lower-level insects are remarkably creative and diverse. Biological research has revealed that some insects can process information from tens of thousands of ommatidia using very small nervous systems. The functions of the individual eyes in a wolf spider's compound eye are not entirely the same; some provide forward vision, while others detect and provide peripheral visual information. These visual signals are then fused in the brain to perform motion detection, distance estimation, and motion control. Therefore, through distributed functional partitioning and appropriate sensor routing, the information processing mechanism of compound eyes can be highly efficient. Inspired by this, WiME, an intelligent environment based on a wireless compound eye network, is established. In this environment, each ommatidia will consist of low-resolution visual sensor nodes, communicating via the IEEE 802.15.4 protocol and organizing to form an environmental nervous system. Biomimetic technologies and algorithms will be explored to support intelligent information processing (including path planning, behavior coordination, sensor fusion, and routing) and visual servoing in the wireless compound eye system. In WiME, each eye provides explicit semantics (location information and behavior set), and their topological connections will be implicitly established using behavior networks; behavior will be triggered by events generated by energy accumulation. The connection between nodes includes conditional probability information for information fusion and routing; the robot's navigation will be based on a sensor topology map with explicit semantics, rather than on an unstructured geographical environment. 2.2 Design goals of WiME (1) Distributed compound eyes based on wireless communication Using wireless sensor networks as the connection mechanism, local visual information with high computational load and large data volume characteristics will be connected in the form of explicit semantic concepts to achieve visual coordination with low communication bandwidth. (2) Routing algorithm based on semantics and information holography This algorithm is an optimal path search algorithm for user queries. The routing algorithm should fully consider the query semantics, use information holographic encoding to compress possible queries and global routing tables, and finally achieve fast optimization. This is also a specific manifestation of the information flow distribution of insect compound eyes. (3) Bio-inspired distributed behavior coordination Since this study distributes numerous motion behaviors throughout the compound eye system, the robot navigation control will face the challenge of how to harmoniously organize and trigger behaviors. This research will be inspired by biological neural networks to explore the action organization methods of spike excitation, in order to realize a behavior control network with goal-driven and timely response characteristics. 2.3 Design of Monocular Neurons in Wireless Compound Eye System ① Basic Motion Detector (EMD). Based on the basic motion detector (EMD) unique to biological vision, a low-resolution and randomly readable CMOS vision sensor is used as the retina and connected to the wireless module to form the system's monocular eye. ② Semantic Extraction of Low-Resolution Images. Real-time image semantic extraction includes: low-resolution feature extraction, including color, texture, and region shape; fusing texture, color, and shape features and providing interpretations; region segmentation and spatial analysis. This part will use down-resolution technology, combined with parallel computing and FPGA implementation, to reduce computational costs and apply the algorithm to distributed visual fusion. ③ Object Tracking and Pattern Recognition. ④ Wireless Sensor Nodes. The node connects to the EMD via external circuitry and communicates with other compound eyes and robots via a receiver using the IEEE 802.15.4 protocol, thus providing a channel for the compound eye's nervous system. Simultaneously, the fusion of compound eye information and the generation of behavioral sequences are also performed on the node. 3. Routing Method Selection All robot navigation requires solving the problem of how the robot learns the path to its destination. In wireless sensor networks, routing for information communication between wireless nodes is also a primary issue. As mentioned earlier, due to fixed geographical information, spatial path planning and information communication routing in WiME can operate in the same way. Therefore, the following section uses path planning to illustrate the selection of a routing storage and query method. In wireless sensor networks, wireless nodes, due to energy constraints, use low-power embedded processors with limited computing power and storage space. WiME is no exception; it generally cannot directly store path information or store map information on the node to calculate the optimal path when needed. Therefore, the following four methods are considered first. Method 1: As a commonly used method, path information across the entire map can be queried. Since the number of rooms *n* is large (assuming *n* is not less than 1000), and the path data is enormous (there are *n(n-1)/2* paths), such a map can be provided by one or more master servers. No single wireless node or a finite number of nearby nodes can meet this storage requirement. A natural approach is to store the global map on a server, with the robot terminal downloading path information from the server when necessary. This is similar to how GPS devices work. Method 2: Establish a wireless communication link to the target point based on the broadcast wireless routing communication protocol used, and use this communication line as a geographical navigation route. Method 3: Utilize the idea of ​​dynamic path planning, where each node stores geographical information within a certain range related to itself and generates optimal path information. Method 4: Each node stores the geographical information and connection relationships of the global node distribution, and collaboratively calculates the optimal path with nearby nodes when needed. This borrows the concept of distributed computing in computer networks. Each method has its advantages and disadvantages. Method 1 is easy to modify; adding or deleting nodes only requires updating the master server. Method 2 does not require prior knowledge of the node's geographical location information; the entire path information is dynamically established and modified. The third method dynamically adjusts the optimal path based on road conditions. Since nodes can observe road information in real time, parameters can be introduced to reflect the current surrounding road conditions, such as the degree of congestion, and thereby dynamically maintain an optimal path table containing the node itself and its neighboring areas. However, all three methods rely on multi-hop communication, requiring significant communication bandwidth and latency to return complete path information, which challenges the robustness of the communication protocol. The fourth method requires relatively less storage, on par with the number of nodes, but real-time distributed computation of optimal paths for multi-node collaboration remains a challenging problem for wireless sensor nodes. After all, current distributed computing is still largely confined to the field of computer networks. Applying distributed computing and the latest grid computing concepts to wireless sensor networks may become the next direction in the embedded systems field. In this WiME design, there is no concept of a host; each wireless mobile node acts as both a host and a router—this is an Ad-Hoc network. Routing methods in Ad-Hoc networks can be divided into two main categories: routing based on routing tables and routing based on on-demand route creation. Due to the massive amount of path data and extremely limited storage space, methods 2, 3, and 4 above all employ a routing approach based on on-demand route establishment. While method 1 provides routing based on a routing table through a server, the limited number of servers is unsuitable for such a large wireless sensor network. Is it truly impossible to store such a global routing table on each wireless node to achieve a true routing table-based routing approach? Considering all factors, this paper proposes the following method—target direction lookup. This is similar to asking for directions on the street; the person tells you which way to go; at the next fork in the road, you have to ask again; eventually, you can successfully reach your destination, but the person asked cannot provide the complete route, only a general direction. In contrast, this method utilizes prior knowledge of relatively fixed indoor geographical information; each node only needs to store its own direction information to the target point, with a storage requirement of only O(n); it also avoids multi-hop communication during lookup and does not add additional communication burden, making it clearly more suitable for the characteristics of wireless sensor networks. Therefore, path lookup in the WiME system adopts this method, and communication routes are also established based on this approach. 4. Bloom Filter 4.1 Storage and Query of Routing Information In reference [3], the authors proposed to implement semantic routing in wireless sensor networks. The specific method is to store a semantic retrieval table in each node. Each point in the retrieval table corresponds to a region classification. Each node has only a limited number of region classifications or "routing possibilities". In this way, when a routing query input containing sufficient semantic information occurs, the node calls its own rule engine, calculates and matches a point in the retrieval table, and obtains the information of the next hop to that region from its corresponding region information. This is similar to the single-step path query method in this design. This design also has such a rule engine, namely the Bloom Filter to be introduced below. The difference is that in this design, there is not one retrieval table, but multiple retrieval tables; the elements in the retrieval table no longer indicate the category of region or route, but indicate whether the input is in the current routing table; and the query input is not abstract semantic information, but geospatial identifiers with clear semantics, such as person name, room number or unit name. As shown below, using Bloom Filters can not only solve the problems of route classification and query, but also further reduce the amount of path information data in resource-constrained wireless sensor nodes. Therefore, in the design of WiME, a counting Bloom Filter is used for each packet to achieve dynamic modification of routing information. The following introduces two types of "rule engines": basic Bloom Filters and counting Bloom Filters. 4.2 Bloom Filter Concept The concept of Bloom Filters was first proposed by B.H. Bloom in 1970. Given a set S containing n elements, each element can be a unique symbol or set of symbols that can be recognized by a computer, such as a name, website address, or a number. We define a vector table v containing m elements, where each element in v is represented using only 1 bit, i.e., each element can only be represented as 0 or 1. Initialize each element in v to 0. Assume there are k independent hash functions H1, ..., Hk, with a mapping range of m. For each element in S, after hash transformation, set the corresponding position in v to 1. To determine whether an element 'a' is in set S, we can perform k hash transformations on it (refer to Figure 1) and check if the corresponding element in v is 1. If all k corresponding elements are 1, then 'a' is in set S. For example, if S represents a URL lookup table, and each element contains an average of 50 ASCII characters, then direct storage would require 400n bits; while using a Bloom Filter, it would require m bits (m and kn are on the same order of magnitude). Since hash function calculations take time, limiting the number of hash functions to a small amount significantly reduces storage space; therefore, this is a method of trading time for space. 4.3 Optimal Bloom Filter Positive False Positive Probability As seen above, the number of elements in the set n, the number of hash functions k, and the length of the vector table m are the three key parameters of a Bloom Filter. In a Bloom Filter, there exists a situation where, although an element does not belong to set S, due to the randomness of the hash functions, the corresponding element in v after all k hash transformations might be 1, thus mistakenly identifying the element as belonging to set S. This situation is called a "false positive". From a probabilistic perspective, positive false positives are always unavoidable. The probability that each element will still be 0 after inserting n elements into the Bloom Filter table is (unless otherwise stated, the hash function is assumed to be uniformly mapped below): 4.4 Counting Bloom Filter During the generation of the Bloom Filter table, it is inevitable that the same position will be mapped to v, which will cause problems when there are additions and deletions. If an element is deleted from the set, all elements in its corresponding Bloom Filter table will change from 1 to 0. Then, other elements mapped to that position will not get the correct result when querying whether they belong to the set; this is called a "false negative". Counting Bloom Filters can solve this problem by changing the representation of each position in the vector table from 1 bit to multiple bits. Thus, each time a position is mapped to during an addition operation, the count for that position is incremented by 1; during a deletion operation, the count for that position is decremented by 1. The number of counting bits determines the maximum countable value. 4.5 Other Improvements to Bloom Filters Besides counting Bloom Filters, many other improved Bloom Filter data structures are being proposed. Reference [6] proposed a compressed Bloom Filter that explores the relationship between m, n and k in non-optimal cases; Reference [7] proposed a domain decay Bloom Filter that addresses the characteristics of flooding queries in wireless sensor networks by proposing a spatial domain decay method. The bits set to 1 in its Bloom Filter vector table will be cleared to 0 with a certain probability as the spatial domain changes. Thus, when the Bloom Filter is decoded, it becomes a count of the number of 1s at the corresponding positions of k hash functions (the larger the number, the greater the probability); Reference [8] proposed a split Bloom Filter that addresses the situation where repeated additions and deletions eventually render the initially designed Bloom Filter table unusable by dividing the main table into multiple sub-tables. Considering all factors, the author still believes that the counting Bloom Filter is simple, easy to use, and has good performance. Although Reference [5] suggested using 4-bit counting, after theoretical analysis and experimental verification of the number of counting bits, the author finally adopted 2-bit counting. This can reduce the probability of reverse false detection caused by repeated additions and deletions to 1.85×10-4. Only one false positive occurs after 5396 iterations of adding and deleting data, which is sufficient for a scale of 1000 nodes. However, discussing this issue is beyond the scope of this paper and will not be elaborated upon here. 5. Conclusion WiME is an intelligent compound eye system implemented using wireless sensor networks, inspired by biological behavior. It attempts to effectively reduce the complexity of intelligent implementation from a biomimetic perspective, improving the robot's mobility; at the same time, it broadens the application scope of robots, enabling inexpensive mobile robots to exhibit excellent mobile intelligence. This is a new theoretical attempt to solve the application difficulties faced by centralized artificial intelligence from another perspective. The robot navigation technology in WiME adopts a single-step direction query method to complete the routing query task in the case of one hop; moreover, it uses Bloom Filter to compress storage, and the space is highly optimized. This makes it possible to complete the task of path and communication routing query system, which involves a large amount of data, in the environment of wireless sensor networks with weak computing power and severely limited resources. It is hoped that the ideas of this design will provide good inspiration for other robot navigation applications.