The Mapping Singularity Is Near

A high falutin’ title to convey how maps are converging and fidelities blending toward a “medium definition” (MD) steady state, to empower humans with machine data, and machines with human intuition.

Over five years into CARMERA’s journey, it has been both fascinating and illuminating to straddle two decades, each a unique era with distinct paradigms — moving from binary to convergent approaches toward map building for mobility.

Within autonomous driving, we started to see hints of this shift about a year ago, with a new “Hierarchy of Needs” taking shape, which we discussed last fall.

Since then, we’ve seen from the industry additional recognition of a related trend, which we’d been sensing for a while: Consumer, automotive and autonomous mapping needs are converging toward a common set of representational frameworks, prompting map integrators and data suppliers to rethink their strategies, while paving a path to accelerated deployment for the rest of the 2020s. Given the widespread corroboration we saw of these hypotheses from leaders across the mobility and mapping spectrum, we decided they’d be worth publishing for an even broader set of eyes. We’ll use some of the slides from our presentations to unpack a bit of if here…

The Recent Past & Present

To start, in the 2010s, the “high definition” (HD) map concept emerged as a mechanism for conveying key priors to aid machine-autonomy decisions. This machine-focused HD concept was presented in distinct — almost binary — contrast from “standard definition” (SD) digital maps created for human use, which were approaching ubiquity. Take an example intersection, first mapped in SD:

Now let’s take a look at the same intersection that we mapped in HD for Level 4 autonomous driving:

As you can see from the visuals and descriptions, there is a nearly 100x step-up in breadth, depth and accuracy to go from traditional SD to HD levels of fidelity.

The [Not-Too-Distant] Future

This bright-line distinction between SD and HD, however, is beginning to blur — the result of machine autonomy becoming more sophisticated and human applications becoming more demanding. A new “MD” standard is emerging. The definition of “MD” is still very much in flux and inherently nebulous given its position in the fidelity continuum. But in general, the power of MD comes from fusing the most critical elements of HD precision and insight, with the unmatched scalability of SD.

Glimpses of our MD thinking can be seen from work we did with Toyota’s TRI-AD unit to assess how closely vehicle cameras alone can yield HD levels of fidelity. One of the requirements was to do so with no initial map at all, having to create an SD map out of vehicle trace data alone. The effectiveness and inherent scalability of such an approach underscores the building of our new Change-as-a-Service offering on top of what we call an MD (“medium definition”) map. This MD layer gave us great efficiency in establishing the scaffolding to detect and localize features, and changes to those features over time. It allows us to leverage existing SD maps, whether proprietary or open-source (e.g., OSM), or as noted above in the Toyota example, create them ourselves from vehicle telemetry if we need to. We can then upgrade the MD data as necessary to update HD vectors in a base map that either we or a partner own.

In the process of doing this, we suspected that with just a little more semantic information in the MD data, the map may be sufficient to meet the continually advancing perception and control capabilities of L2 autonomous driving stacks in the near term, and L4 in the longer term. So we generated a hypothetical representation of that same intersection in a future MD state to illustrate:

Note that while the feature data is pared down, it is still information-rich. For each traffic signal, for example, the map describes the order of control, the lane applicability of control and so on.

And indeed, it mirrored the sentiments we heard from companies across the L2 – L4 spectrum:

  • We don’t need every part of the map to be at HD levels of granularity or spatial accuracy like we used to require…

The simplified (work-in-progress) synthesis of all of this is as follows:

What Does This Mean?

If these trends we’re seeing continue, there are a number of implications for stakeholders to consider to ensure they remain future-proofed:

  • Auto OEMs: Traditional HD map building no longer should be an impediment to greatly expanding L2+ hands-off operating domains for increasingly discerning car buyers. One of the obvious responses to this might be: Isn’t Tesla already doing this without maps at all? The answer is — not quite. While they have spoken publicly about their aversion toward “HD maps,” Teslas today do use higher definition data than found in a conventional SD map (e.g., more information on things like lane counts, turning options, traffic control). While many find their approach to “Full–Self-Driving” problematic (including the term itself), this leveraging of enhanced map information is useful to understand what is (and isn’t) possible with ~MD today.

Thanks for reading, and do reach out with any views that might further our collective thinking.


Special thanks to co-authors/editors, Justin Day and Ethan Sorrelgreen, and to expert reviewers, Oliver Cameron, Muthu Kumar and Marc Prioleau.




Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store