Thread by @wintonARK

Perhaps useful (at this point) to compare to Tesla Vision to HD-mapping strategies where an understanding of the road layout and intersection shapes, stop-signs and traffic flow lane-mappings are held in-memory.

🧵

https://twitter.com/wintonark/status/1428828394150502403

One can simultaneously understand why a Waymo or Cruise would pursue the HD-mapping strategy, and the vulnerabilities of such a rigid approach.

/1

Intuitively, you and I drive more easily on roads that we already know;

holding HD-maps in memory narrows the problem-space—the dynamic objects are things that don't appear on this map--in a way that appears more tractable.

/2

One can also understand how HD-mapping might be seen as the right minimum-viable-product pathway to market.

/3

Yes, for each additional terrain you attempt to launch you have to HD map and label it, so there are scaling issues, but this should allow you to demonstrate proof of concept in a single service-area sufficient to justify capital-outlays required to scale to other metros

/4

BUT when you or I encounter an uncertain situation on a typically familiar road we default to a more generalized understanding of the driving surface and intended throughways. These situations (road construction, lane re-mappings, &c) are more frequent than we appreciate.
/5

If a system has been designed to use that HD map as a crutch and so has a less well-developed generalized model for drivable surface and dynamic objects, then it is likely to catastropically fail just when it needs to be most robust (traffic cones! on the road! 😱)
/6

In effect, in the conditions that are most likely to produce the corner cases that result in accidents or imprecise driving, the HD-maps-based system's performance deviates below a generalized vision models performance.

/7

Net, an HD-maps based system may appear better for the boring parts, but then become worse for the parts that matter. (You think you are getting closer to commercialization but you are in fact resigning yourself to underperformance when performance is most critical.)

/8

Couple that with the expensive sensor-suites that HD-maps based companies have converged upon, the attendant computational load that those suites require, and the data ingestion constraints that commercially imposes (since you can't install the gear on consumer vehicles...)

/9

One can understand how biasing the system to more generalizable and scalable solutions early in the autonomous perception-prediction-planning pipeline may yield an ultimately more robust foundation onto which to build the rest of the stack.

/10