Module 6: National Contest High School

Module 6 – High School

This module has programming notebooks using data from space sources to examine how space data can be used and displayed for decision-making.

Part 1 – Optimizing data transfer and decision making using high quality labeled data

Part 2 – Contest type problems

With the explosive growth of artificial intelligence paradigms, a wide array of applications can be found in space missions. In particular, robotic missions on Mars, such as those carried out by NASA’s Jet Propulsion Laboratory in the past decade, have provided a window of opportunity for the collection of invaluable data for machine learning and deep learning algorithms. Consequently, potential tasks for these algorithms include aiding navigation through treacherous terrain and facilitating the use of numerical data for activating some functionality during a mission.

The importance of optimizing data transfer and decision-making during a rover’s mission cannot be understated. Current difficulties center around the delays in communication that occur between the Deep Space Network (DSN) on Earth and a given rover on Mars. While communication with orbiters that relay critical information back to Earth is adequate for ensuring that information arrives within twenty minutes, it certainly leaves much to be desired due to the constant influx of information that arrives from the Martian rovers. Given these limitations, machine learning algorithms could play a crucial role in allowing these rovers to make decisions in a more independent manner that does not rely on a series of potentially delayed decisions by controllers back on Earth. It is important to emphasize, however, that the robotic systems of the rover are not equivalent to the artificial intelligence algorithms that are being discussed here. The artificial intelligence models make use of the rover’s equipment, such as its cameras and sensors, to enhance its capabilities through the collection of data in different modalities for training when given a particular task.

Despite the promising advantages of the use of machine learning algorithms, their success is driven by the availability of high-quality and often labeled data. Think about a task that we mentioned above. What modality of data would such a task involve? Will the task need data that is labeled; in other words, is the task a supervised learning task? Is the task a classification or regression task? Is deep learning or classical machine learning a feasible approach towards ultimately facilitating the rover’s operations under the assumption that the task is successfully completed? These are all important questions that you should be asking yourself when faced with a problem that will likely involve artificial intelligence algorithms.

CONTEST TYPE PROBLEMS

To begin training yourself to solve problems related to the topic at hand with machine learning models, consider the following task. You are already aware that there exists a relatively acute disparity in time when communicating with a given rover. When gathering important information, these rovers will often be sent on missions that involve navigation. How can we supplement the decision-making process that is involved when deciding how to best optimize a specific trajectory when faced with uneven terrain? The objective is to maximize the efficient use of the rover’s energy systems. As such, we could consider a task whereby the rover’s cameras and sensors would be involved in not only understanding the various terrain features that it sees, but also making use of atmospheric Martian data for additional independent decisions. As such, this task would involve both an object detection algorithm and a classification algorithm for structured, tabular data. The deep learning algorithm would be fed an assortment of images of Martian terrain that were ideally captured from the rover’s perspective. These images would be labeled with bounding boxes that specify where a potentially dangerous or interesting terrain feature would be located in the camera feed. When a detected terrain feature is deemed to be dangerous or too difficult to traverse, it would allow the rover to take a path that would circumvent the potential nuisance. At the same time, however, the rover must also account for current atmospheric conditions to determine whether its temperature control systems should be activated. These systems are indispensable for ensuring that the rover’s internal components remain intact from exposure to relatively brutal atmospheric conditions. With the use of a classification algorithm, however, we could immediately predict whether the rover should activate the systems or not when given a specific data sample of atmospheric conditions. Ultimately, your task would involve both image and structured (tabular) data modalities that require labeling and training machine learning and deep learning algorithms. Thankfully, however, you would already have existing datasets with labels available to you with which you will be working. Additional data that is relevant and high in quality is always welcomed by machine learning algorithms, but the base data should be sufficient for the task at hand. You would, however, be responsible for any data augmentation schemes and using appropriate models such that your evaluation metrics are as adequate as possible. Consider situations where some training examples may also need to be either modified or removed due to possible lack of quality.

An example of how data for this particular task is labeled for object detection is shown below. Think about why specific parts of the images are surrounded by bounding boxes.

Credit: NASA/JPL-Caltech