Skip to main content


about Pei Zhang

Pei ZhangPei Zhang is an assistant research professor at the INI and ECE. Dr. Zhang's interests include distributed sensing and processing in/between these heterogeneous systems. He obtained his PhD from Princeton University under Prof. Margaret Martonosi. As a part of his dissertation, "Collaboration and Adaptation for the Longevity of Mobile Delay-Tolerant Sensor Systems", he developed a sparse mobile sensor network to track wild zebras in central Kenya (ZebraNet). This project was the first to explore sensor networks in a sparse and mobile setting.

[ email ] | [profile]

CyLab Chronicles

Q&A with Pei Zhang

posted by Richard Power

CyLab Chronicles: Most current wireless sensors networks are static. What are the limitations of this approach, and how do these limitations impact the usefulness of such networks?

ZHANG: Sensor Networks are used in many monitoring applications including military, industrial, scientific, and consumer. Because of their low cost, a wide area can be monitored with relatively low device cost. But limitations stem from their static or nature, including:

  • Deployment logistics - The complexity and cost of computing efficient placements and manually deploying nodes in large numbers.
  • Environmental adaptation - The difficulty in ensuring network resilience in dynamic environmental conditions.
  • Infrastructural dependence - The need for large-scale infrastructure to continuously support sensor nodes that monitor localized but transient events.
  • Maintenance - The impracticability of replacing faulty or energy-deficient nodes in large-scale deployments.

These limitations severely hamper the usefulness of sensor networks for many applications, especially in emergency scenarios such as after an earthquake or during a fire. In such situations, individuals may be trapped in enclosed spaces or buildings, which may be unsafe for rescue workers to enter. Sensor networks can be used in such situations to detect survivors and alert rescuers of their location. In these situations, it is impossible to safely deploy a traditional network.

In addition, a very large number of sensors are needed to cover all affected areas simultaneously, leading to great expense in terms of hardware and deployment costs. While some approaches alleviate the problem by attaching sensor nodes embedded on people, it gives no assurance of adequate node density and does not guarantee that physical configurations will be adequate to route information back to search teams.

Pei Zhang and student Aveek Purohit (pictured) are working to make SensorFly a reality

CyLab Chronicles: What is SensorFly? What is unique in your approach? What problems does it solve in the real world?

ZHANG: SensorFly is a controlled-mobile flying sensor network platform. To the best of our knowledge, it is the most lightweight flying sensor platform implemented to date. SensorFly, with its miniature helicopter-based mobile sensors, addresses the shortcomings of the static sensor networks approach.
In particular:

  • The small size of the sensor nodes facilitates maximum access to the environment.
  • The relatively low cost permits deployment in much larger numbers.
  • Since they are controlled-mobile, the nodes can scout and self-deploy, guaranteeing adequate sensing coverage and maintaining network connectivity.

Mobile solutions along these lines have been explored using approaches from robotics research. Generally research focuses on monolithic devices or a small team of remote-controlled or autonomous robots. Research efforts to evaluate such devices typically focus on individual stability and independent navigation. This focus requires many sophisticated sensors and high processing needs, which results in large and costly devices. With SensorFly, we drastically reduced the number of navigational and control sensors compared to other flying platforms to include only an accelerometer and a compass. This significantly reduces weight and allows for a much smaller system size, but it increases the difficulty of many system functions. We introduce novel approaches to make such a system feasible.

In the real world, such a system would mean drastically lower cost in emergency applications. Reduced cost in terms of physical system cost due to reduced nodes, and cost in terms of maintenance and deployment of such systems.

CyLab Chronicles: Tell us a bit about your use of autonomous mapping, localization and navigation services, and in what ways these features offer significant improvements over static sensor networks in terms of deployment, sensing and maintenance?

ZHANG: In any mobile system, the devices need to figure out where they are and have a sense of their surroundings and SensorFly is no exception. In robotics, simultaneous localization and mapping techniques are usually based on multiple accurate and multiple range sensors (localization), along with dead reckoning (tracking device movement). Not only is the SensorFly system highly constrained in terms of sensors, dead reckoning is inaccurate for the flying platform. Our protocols supply these services with a minimum number of inexpensive, versatile and lightweight components: the radio, the accelerometer and the magnetic direction sensor.

The constraints inherent in a platform with our form factor make it infeasible to obtain highly accurate localization and mapping. However, we observe that such accuracy is not always needed in sensor networks. The system can be made viable if the metrics are “good enough.” The SensorFly system is designed to be resilient to large input errors. For example, the system assumes inaccurate maps and location estimates. As the nodes themselves are lightweight and resilient to collisions, the navigation protocol adopts only a best-effort approach. It is advantageous, but not essential, to avoid unnecessary collisions with obstacles.

To work around our limitations of sensing, we developed a protocol called “Hopscotch” for localizing nodes. The overall movement pattern resembles a game of hopscotch. A typical group consists of 4 nodes, with 3 acting as anchors to provide localization at any given time, while the 4th node is mobile. Within a group, nodes keep changing roles to move around the environment. From our radio characteristics, we observe that scout nodes obtain better location estimates if they are within a favorable distance (about 50cm) from each anchor node. We keep the nodes constantly within this range for better precision. If the nodes for too far, errors might accumulate, therefore, at the same time, certain beacon nodes or anchors may be deployed. These nodes act as location markers and, when near obstacles, to alert approaching nodes. The system utilizes this information to localize and map their environment. Thus, by combining redundant, inaccurate but inexpensive methods, the Sensor-Fly system can achieve sufficiently good performance.

CyLab Chronicles: What particular challenges and constraints have you come across in the development of SensorFly?

ZHANG: Working on such a novel platform certainly brings on many challenges. Most work has either used low capability sensor nodes, or highly capable robotics platforms. Therefore we had to construct the SensorFly platform before doing any minimal sensing research. Borrowing mechanical drive components from off-the-shelf toy helicopters, the micro-mosquito, we were able to drastically reduce the development time of our sensors. This combined with the custom electronics system we were able to design and implement such a new system in a short amount of time.

Through the design process, one major challenge was meeting the weight constraint of our system. Like many battery-operated sensing systems, we have the energy and cost constraints. In addition to traditional sensor network constraints such as low per-device cost and energy usage, the helicopter platform requires that we consider a new metric: weight. The entire craft, including the battery and driving mechanisms, has a target weight of 30 grams, the weight of approximately 10 pennies. With our current drive mechanism, a node that is heavier than 30 grams can still fly, but it exhibits elevated energy consumption parameters. The absolute maximum takeoff weight is only slightly higher at 35 grams. This weight constraint also limits the number of sensors that can be carried by each node. It is impossible to incorporate a full array of sensors on the sensor nodes and still maintain reasonable performance. Therefore, each node must work with others in order to collaboratively maintain position and monitor the environment.


CyLab Chronicles: What commercial applications do you see SensorFly contributing to?

ZHANG: The SensorFly system can be applied to many applications where autonomous and rapid operations are needed, for example, during emergency rescue operations. During these operations, the goal of rescuers is to save as many lives as possible without needlessly risking their own. However, structures that remain standing may be inherently unstable and risky. Without appropriate guidance, the rescue workers must assume grave risks to enter and search unknown areas in search of survivors. At the same time, survivors often remain trapped in enclosed spaces within buildings. In many cases, survivors are not conscious or are immobilized and therefore unable to attract the attention of rescue workers outside the building. In the face of hazards like smoke, heat and open electric lines, the survival rate of individuals drops steeply as the time that they remain trapped increases. SensorFly can enter the building in this scenario and search for survivors without risking lives or wasting the time of the rescuers.

Thus, the SensorFly system could benefit applications that needs:

  • Rapidly and efficiently search unknown areas
  • Locate and search objects and guide rescue workers to these objects
  • Be resilient to dynamic environmental conditions and system failures with a high degree of autonomy.

We foresee this system naturally fitting into applications such as:

  • Emergency response applications such as an earthquake or a fire
  • Military applications such as urban combat situations.
  • Security alarm applications, such as home intruder scenarios
  • Elder care, such as in-home emergency verification

See all CyLab Chronicles articles