To begin our design, a task analysis was done examining the typical user and what steps they might take to walk from one destination to another. This analysis was completed with main tasks as well as sub tasks. Taking it further, our team began a task classification by deciding what types of information were needed for ever user action and decision. Finally, we landed upon an automation strategy based on a thorough task allocation study and subsequent design of the levels of automation. 

The Nite Knight system will have a companion app that allows the user to interact with and set up their robot. This app will also allow Guardians to keep track of their loved ones using Nite Knight by feeding them location data and providing insight into their walk. A Guardian is anyone designated by the user to be allowed to access their Nite Knight information - typically loved ones, parents, etc. The video below gives an idea of what the Guardian experience on a mobile device might look like. 

Threat detection by Nite Night will take place with a trained AI. As with all AI models, the Nite Knight AI will need a set of quality training data in order to recognize and accurately identify threats. This training can be broken down based on our available inputs: video data and audio data. For video training, our designers need to make sure to train the AI with sufficient amounts of data on each harm the system should detect. For example, training videos must include large amounts of data on aggressive humans as well as aggressive animals. Audible training will be similar, where each hazard or alert noise will need to be trained sufficiently with enough examples of each.  

The main body of information for aggressive creature identification will be police body cameras. This source provides a consolidated place to find video of public danger, without having to sort through unfiltered internet videos or create large amounts of training data from scratch. As our device is aimed to be used in higher population density areas, this training data will be ideal as it includes all types of living creature-based hazards our users might encounter. This data will need to be curated to avoid bias shown by law enforcement, and the AI will need adjustments and testing to confirm there is no bias in threat detection. This topic is addressed below in the ethical implications section.  

In addition to body camera footage, our team will incorporate targeted video training based on research into body language indicative of aggressive behavior. With Google AI’s Teachable Machine, our team was able to generate a demonstration of how this might work. With a training set of data aimed at identifying several different poses, the AI model can identify aggressive poses in real time. The video below demonstrates this AI's functionality. It was trained using several short videos of the operator performing different friendly and aggressive stances. 

Karen age 70 retired but still lives a lively and active life and enjoys taking evening walks. Recently Karen has been fearful of walking at night especially since her partner passed away recently and she now lives alone. Her children live a few towns over. She and her children recently decided to purchase a Nite Knight in hopes of allowing Karen to go on walks independently at night with peace of mind her family would be alerted if an emergency were to occur. The follwing chart is the process Karen goes through to use the Nite Knight for the first time.

Ethical, Societal, and Safety Implications  

An AI model can only be as good as the data that is fed to it. Our society contains bias and prejudice, and without proper care and consideration of this in our training data, it could make its way into the AI model. If our AI model, on an average, considered certain races or ethnicities “more dangerous”, our product would be a failure. To avoid this, in-depth research and focus on universal signs of aggression and danger would be the only data considered for the AI model.  

Our data aims to make individuals feel safer, but to do so, we use features that could cause risk to the user if fallen into the wrong hands. Learning from the issues with Apple’s Air Tag product, our location tracking will have the highest data security possible to ensure that user’s location data is only accessible to those chosen by the user.  

In discussing our product, our team felt strongly that this robot's design should not ostracize the user and make them feel embarrassed. The last thing our team wants is to create a product that people are ashamed to use. To tackle this, our robot has several modes varying in interaction level. Using our machine in the lowest interaction mode allows users to interact with others without coming off as unfriendly or strange. In this mode, the robot keeps itself quiet and small, following the user without obstructing the line of sight to strangers.  

Limitations

While machine learning has recently rocketed into the mainstream, it is still in its infancy. Powerful chat bots like ChatGPT and Jasper are becoming more powerful and knowledgeable every day, they still fall short of what human thought can produce. For Nite Knight, this implies that our danger detection AI may need several AI advances before it becomes feasible. While we demonstrated that a simplistic AI can detect obvious body language, it is unclear what technological improvements would be necessary to increase this resolution. The Nite Knight AI needs to be correct to high confidence when it alerts that there is danger. While the system can have safeguards to prevent false alarms, the mere presence of false alarms can erode confidence in the system.  

Nite Knight signifies a pioneering leap in automation, marking our entrance into the security product market. Selling a first-to-market product like this comes with its own set of challenges, particularly when it introduces the concept of incorporating an escort robot into daily life, which might not have been on the public's radar. Our team recognizes the significance of ensuring a seamless and inconspicuous rollout for Nite Knight. We've seen instances in other cities where attempts to introduce automated security robots have faced obstacles, leading to prematurely terminated pilot programs. For example, in 2021, the NYPD introduced Digidog, designed to investigate high-profile hostage areas, but unfortunately, it only lasted a mere 3 months due to a public outcry. The outcome highlights the need for a well-thought-out approach. When introducing Nite Knight, we'll focus on understanding the scope of the release, potentially starting with a controlled pilot program in select small neighborhoods before embarking on a broader rollout. 

To achieve the full integration of Nite Knight into society, we'll need support not only from customers but also from the city administration and law enforcement agencies. This aspect poses significant challenges, particularly given the current heavy regulation of autonomous robots in specific regions. However, it's essential for the project's long-term success and the safety of our customers that they're aware of the robot's connection to the police grid. In case of emergencies, the robot can promptly contact the necessary officers and authorities. Our vision includes a small-scale pilot program lasting several months, aiming to demonstrate the tangible benefits of enhanced safety when walking alongside our robot. A successful pilot could pave the way for a smoother process in convincing the city to embrace Nite Knight as a valuable assisting entity. 

Future Directions