Perfect Recall: Storage and Robotics

Storage may not be the sexiest megatrend in robotics, but it is critical to outperforming and out-adapting your competitors.

Earlier this month, I wrote about how advances in sensors are allowing robots to understand their environments significantly better, but that’s only half of the story. Robots need to remember what they’ve seen before, and better data storage and analysis allow robots to compare and contrast their futures with their pasts.

The drop in price per gigabyte of storage has led to tremendous opportunities. Over the last few decades, the price of storage has dropped nearly seven orders of magnitude. As Matthew Komoroski, Lead Software Architect at TROVE, has nicely shown [in the chart below], it’s cheaper now than ever to store data. An amount of storage that costs a few dollars today would have cost so much in the 1980s that it would have been a C-Suite decision. (Mind you, this chart is on a log scale.)


Today, you can collect and store all the data an application needs and not have to make dramatic cuts to capability. That ability to retain all the data a robot has ever collected results in perfect recall. Multiplied across a fleet of robots, this means that the first time a single robot finds an edge case, it can be solved and propagated to the entire network.

As always for these types of problems, you need more data to find the rare edge cases. Driving is a great example. If something only occurs once in a million driven miles, it seems rare. But according to the US Department of Energy, there were 3.17 trillion (with a “t”) miles driven in the United States in 2016. So that one-in-a-million instance will occur several million times in a given year. Therefore, if you only have the data from the first 10 million driven miles, you might only have a handful of examples of this event. If you couldn’t store all the data ever collected, you would have zero chance of seeing it enough times to solve it.

Without these advancements, our robots would be like the main character from the movie “Memento,” unaware of what happened in the past, making the same mistakes over and over.

The key is the ability to analyze this data and then create algorithms and standards for what to do when the same situation occurs again. This, of course, is the mostly over-hyped machine learning aspect of robotics today. The reality of data storage involves a lot of manual tagging and custom algorithm development for each case.

Almost every robotics company has small teams of people who just wade through the data and try to understand what the robot didn’t recognize. They then develop a series of tools to try to create more of that recognition automatically. Over time, this leads to a more robust system that needs less intervention.

For our portfolio companies like Dishcraft and Marble, this becomes part of their competitive advantage. With every plate that Dishcraft cleans, they get data on how to clean the next one better. For Marble, with every block they drive, they are in a better position to navigate the next one. Over time, this gives these first-to-market companies an edge because their rivals will take years to collect the same amount of data. In that time, our guys will be even further ahead.

Relatively inexpensive perfect data recall for any robotics enterprise, from a two-person startup to a Fortune 500 company, is a massive game changer. If you’re working on robotics applications that are just now possible due to the availability of data storage, please let me know @nomadicnerd.