From Recall to 2nd Generation Device
Inevitably hardware development and prototyping hits a wall sooner or later. We just hit one, and it will probably not be our last wall either. Today we issued a recall on ALL the devices deployed in Austin in February. The reason: partly structural and partly code related.
On the structural side, we found that the gap which exists between our device and the bin walls had an impact on the weight captured — up to 25% error margin was detected. Looking at it in hindsight, of course this makes sense. The trash bag is protruding over the device top plate and waste is getting into the perimeter gaps, and depending on how much waste gets into the gaps determines how much of the weight is lost to gravity rather than captured by our device.
To understand what’s happening you can try this simple experiment at home. Position your personal scale 30 cm off the ground, then put a bag full of beans on the scale. Let the bag (and beans in the bag) protrude off of all the edges. Record the weight as value A. Next, take the same bag with the beans and put them both into a bowl. Place the bowl on the scale. Read the measurement — record this as value B. Note the difference in value between A and B; that’s the error margin caused by the gap, and the problem that destroyed our first datasets.
Why the gap? It was a design decision. When we initially built the device we presumed the install would be under the bin bucket, but we quickly came to the conclusion that because of so many different bin setups and home kitchen designs, there was no way we could mount the device under the bin. It would have to go into the bin bucket itself. Fine, we solved one issue. But we created another one, the gap. To be able to fit our device into each bin we asked our customers for bin measurements — internal bin dimension (L x W) in inches. Seems simple enough, but much to our surprise, we realized quickly — within our first two installs — that our instructions on what measurements we needed were not precise enough; we failed to specify top vs bottom and internal vs external (L x W). It matters.
Our devices didn’t fit — we weren’t even close. The traumatic experience of having to cut the top plate on the spot, including the skeptical looks from our first customers, informed our next set of decisions: A) We were much more specific as to what measurements we wanted from our customers, and B) We took off an inch from each end of the top plate, just to be extra safe.
In the second round of installs the devices fit. No problem. But, as pointed out, now there was a gap, which we proceeded to ignore much to our own detriment. The gap, as it turns out, made our data from Phase 1 impossible to trust. Which brings me to today, and the recall. At least now we have a solution for the gap issue — mounting an adaptor piece (cut to spec for each bin size) onto the existing top plate.
The second problem we discovered and which needed to be addressed is the way our first prototype is measuring and aggregating the weight of household waste. The sensor was wired and configured to record every motion (including pressure) as true weight. This means we were also capturing the act of the human hand pressing down on waste as weight added. In our datasets we could see sudden spikes in weight, then self-adjusting in gradual declines. Intellectually, this makes all sense, of course. When a hand pushes on the bin, it exerts force, and that force equals weight. The hand then releases, kicking off the process of recalibration. More surprising was the fact that this process of recalibration can take up to 4 hours. We only found this out during troubleshooting, and through reproducing the full data log from one of the households in our garage.
But more importantly, each calibration window didn’t last the same amount of time. Rather, it seem to depend on how much pressure was exerted, and on the design of the bin. Because of these differences the data processing and plotting became tricky for us, and unsystematic. We were deciding on a case by case basis which peaks were to be cut out and which data points we would treat as true data. We lost trust in our own data.
Rest easy, we now have a solution for this issue — a delay function of up to 30 seconds between the time the sensor registers a change in weight and when it records it. Going forward, if there is human interference involved (pushing down on the waste or the wrong item is discarded into the bin by accident) the customer is able to correct course without the device capturing human error as true waste.
Throughout the process of troubleshooting we also managed to connect our entire system to the Cloud which now makes possible remote access to each device and each display unit, thus giving us the ability to actively monitor and interpret customer data, reset the display unit remotely, and supply customers with notifications in waste spikes and progress made in meeting individual reduction targets.
I won’t deny it: we were staring at a blank wall last week. Solutions weren’t coming. The anxiety that comes with each new failure of the device is real. And yet, it’s worth reminding myself, my team, and other fellow entrepreneurs, developers and engineers, that in a moment of despair and mounting problems, focus on the task at hand, solve one problem at a time, and keep building. The solution will come.
On our end, we are scheduling device pick-ups in all ATX households involved in our pilot project. The devices will be upgraded with the newly outlined hardware and software solutions, and redeployed as BINIT Second Generation units back to our customers for the start of Phase 1 by early April.
Binit, building pathways to understanding your waste.