As a Master's capstone project for the program at UW, our team partnered with and their Sensing, Inference & Research Team. Uber Eats currently uses sensor and GPS data to track the delivery process, but was interested in understanding how Machine Learning could help to identify and ultimately prevent delivery errors. Uber Eats knew they had a problem—and had a new tool they wanted to use—but lacked contextual awareness of the errors that couriers experience. UBER wanted to focus on the "last 50 meters" of the delivery where they felt the majority of errors occured. These errors not only hindered delivery efficiency, but also made for a poor user experience. To tackle this problem, we needed to discover the problems and errors encountered by UBER Eats couriers.
Early on we ran into a difficult problem. UBER was unable to provide us with screenshots, wireframes or user flows of their app.
To surmount this, each team member signed up to be an UBER Eats driver and one of us actually completed several deliveries, taking screenshots at every step of the process. Using these screenshots we then rebuilt the user flow to get a holistic understanding of the courier app. We drafted a detailed research protocol and began recruiting through online communities and in person outside busy restaurants. Our users came from three distinct groups: couriers, eaters and restaurant staff. With more than 80 responses to our screener surveys, we were quickly able to begin analyzing qualitative data about the food delivery experience.
We conducted semi-structured interviews with five couriers, four eaters, and four restaurants. Users walked us through their process from start to finish, recalling various problems they had encountered. In addition to the courier interviews, we also conducted two ride-alongs with couriers, observing and recording their process in detail and asking questions to understand why they did what they did. We were able to uncover several problems for each distinct user group, but we identified the courier experience as having the greatest opportunity for positive change.
We learned a great deal from speaking with couriers, from the frustrations they had with UBER support, to a lack of trust regarding UBER's business practices, to finding parking, or even communicating with restaurant staff or inaccurate wait times and restaurant information in the app. UBER Eats couriers are fiercely independent and incredibly resilient. Most of the time, when a courier experiences an error, they figure out a way around it on their own. Rarely do they ever report these bugs, due in part to a lack of reporting tools, but also a lack of confidence in support staff, and the cumbersome and time consuming nature of the current support system. UBER Eats couriers care about efficiency; making as much money as possible in the shortest amount of time. Time spent on the phone with support affects their bottom line. With that in mind, we also had to be conscious of the fact that changes to the courier app affects not only their experience, but also their livelihoods.
Perhaps the biggest takeaway from our research was that the majority of problems couriers face aren't in the "last 50 meters" like Uber had identified. Occasionally they do miss a delivery, but their frustrations are centered around the process of picking up orders and the experience they have with restaurants. We presented our findings to UBER and they agreed to let us focus on the food pick-up experience.
We designed or re-designed three features within the UBER Eats courier app.
- ● A new parking overlay feature.
- ● A re-designed order pick-up flow.
- ● A new contextual wait-time questionnairre.
The goal of the parking overlay feature is to provide couriers with probability based parking locations. UBER is able to track previous pick-ups from a location and use that data to drive an heatmap style overlay of likely parking locations.
The order pick-up flow was reimagined, first adding in restaurant contact information, images of the pick-up location and any restaurant-specific instructions. The order detail screen was redesigned to highlight the most important information to couriers, and to help minimize language barriers. Customer names and order numbers were enlarged and placed at the top of the screen. Food and drinks were separated and highlighted with icons. A section was created to surface special instructions, which were normally only visible to restaurant staff. The traditional 'Next' button text was replaced with "Got Everything" as a further reminder to verify the contents of the order.
Often orders are not ready when the courier arrives, causing delays of 5-15 minutes. Our solution uses that wait time to inform other users by asking simple questions about a location. This allows couriers to share realtime information with UBER and other couriers about their current experiences. This 'humans-in-the-loop' approach is an attempt to feed data into Uber's machine learning algorithims to ultimatly improve the couriers trip efficiency through new features or designs in the future. Collecting such data could help a courier reduce their trip time, helping them maximize earnings and feel less frustrated with the delivery experince.
We recruited six UBER eats couriers to participate in a series of 45 minute usability testing sessions to evaluate our prototype. Their feedback was invaluable and allowed us to address a number of problems with our prototype, from not understanding the heat-map style overlay, to unclear interactions on the order screen.
Using the feedback from our users, we redsigned several aspects of our features. The parking overlay was removed and added directly to the base map. The heat-map was simplified to simple colored shapes corresponding to potential parking, and any designated restaurant parking was labeled.
The restaurant info screen was removed and the pick-up button added to the map screen. On the order detail page, information was further simplified, and a new 'Waiting?' toast notificaion on the edge of the screen guides idle couriers to the questionnairre feature.
The questionnairre proved to be a popular feature sice it allowed useful insights about locations to be shared with other drivers by answering easy questions about their experiences. Answering questions about the restaurant context that only a human would know (vs. sensor data) helps improve the accuracy of Uber's machine learning algorithms that will in the end help improve couriers experiences.