New Roomba j7 with front facing camera can avoid obstacles and takes a picture of objects it avoided. We developed an image review feature where user takes an action based on the obstacle pictures the robot has taken to improve cleaning performance for not only themselves but for everyone in the community collectively.
ROLE
Lead Designer
TIMELINE
Jan 2021 - June 2021
Although our robots were capable of avoiding obstacles, user’s input was necessary to improve cleaning outcome and sharing the images to the system would facilitate learning process and make robots smarter. There were already competitor robots with camera that showed pictures of obstacles it avoided while cleaning but our market research discovered that these robots were not meeting the user’s expectations, disincentivizing engagement.
Users had very little control over what they could do with avoided areas. They could just view the pictures and locate them on the map or re-clean every missed spot without granular control. Also, there were lots of mis-categorized obstacles but users had no way of improving the robot’s detection intelligence.
People were concerned about the data privacy on how the pictures are used or shared. They were likely to react based on their perceived image of the brand and trusted iRobot, but wanted clear message or confirmation on data control.
Offer actionable options that immediately impacts cleaning outcome. This will drive users to review more obstacle pictures, making robots smarter and clean better in upcoming tasks.
Make sure users understand how the data are used and they have full control over the data. They have to trust us in the first place to opt-in to share their data with us and collectively improve robot intelligence.
Sometimes, it’s really hard to understand why Roomba skipped certain area even with the picture. We highlighted the obstacle that Roomba detected during the cleaning job and indicated where the event occurred so that they can have better understanding of robot’s behavior.
Users can decide what they want to do with avoided areas individually with a single tap.
We strongly suggest that users contribute their data as it will hugely improve robot intelligence and cleaning behavior for future cleaning jobs. But users have full control over their own data. They can select specific pictures to contribute to the database or none. They can even choose not to take obstacle pictures during an entire mission.
Users don’t have to wait until the next cleaning job to clean skipped areas. After reviewing all the pictures, users have an option to tidy up those areas to fully clean their home. Pictures which users selected to place a Keep Out Zone will not be shown as the Roomba won’t enter those areas anymore.
We have successfully launched the Roomba j7 with image review feature and accomplished the 2 main goals we set for this project.
Users who have reviewed images after mission saw significant increase in their mission completion rate resulting in cleaner home.
+21%
Increase in mission completion rate after image review
We initially set our KPI for rate of images contributed to database to be 30%. After launch, we saw 38.9% of images being contributed to our database, helping the algorithm accuracy. Users not only saw the value in sharing images but also trusted us to use their data.
+29%
More reviewed images contributed to database vs intial KPI
We initially focused on finding user expectations on reviewing images. We wanted to understand what images are considered sensitive or private and how they expect the data to be handled by iRobot. And what users would expect from the autonomous robot taking pictures of your home especially when there’s a false positive. Here are some of the example pictures taken from prototype robot that were presented to the participants.
- User research participant
Most participants did not express any privacy concerns about seeing the images within the app. It was only when discussing whether to share images that they expressed concern about their data.
Images that felt the most impactful on the robot’s performance had the clearest value to participants, and they were most comfortable sharing these images. Value included a visible hazard or a clear false positive that others might experience.
Participant interpretations of false positive behaviors varied based on previous experience with RVCs; participants expected the app to provide information and control for the behavior they saw.
We explored many potential review flows to surface the obstacle pictures and get feedback from users. We brainstormed what actions users could take at the moment of review as it depended on the user’s physical location. After exploration, we tested our designs to collect feedback and iterate. There were especially helpful feedback from beta testing as the most of the user experience was as intended. We updated small but important details before the launch. We made copy updates to make things crystal clear and align user expectation to actual robot behavior.
We had landed on these 4 choices for users to select but had gone through many copy updates to correctly communicate the result of each choice. First version represented the actions the robot would take but it wasn’t clear what ‘avoid this time’ meant whether obstacle was true or not. Also the options sounded like the choices were about the robot’s behavior around the obstacles but they were actually for the areas. Ultimately we focused our choices on evaluating object detection and removed robot behavior nuances.
Because we heard positive feedback on the idea of tidy up mission, we naturally included the tidy up as the last step of the review process. But to the users, it was a separate action. When users started reviewing pictures, they weren’t expecting a decision to run tidy up mission at the end. Thus we broke the flow into two parts, one for review and one for tidy up mission as an option.
