Against The Odds: Why The AV Industry Should Probably Focus On Just The Even-Numbered SAE Autonomy Levels
TL;DR…
- As the autonomous vehicle (AV) industry evolves, the Society for Automotive Engineers’ (SAE) levels of autonomy have become increasingly debated, and in some cases, can encourage undesirable outcomes.
- With a lot more mileage, driver, rider and stakeholder feedback under our belt, now may be the time to prioritize the levels of autonomy that should exist based on the safety, technical and commercial realities we now know to be true.
- Avoiding the odd-numbered SAE levels appears to be a useful heuristic: Level 1 doesn’t do enough to save lives. Level 3 tends to encroach upon a dangerous tipping point on human attention. True Level 5 either need not or will not exist in a meaningful timeframe.
- On the flip side, Level 0 is what it is and may soon be seen as a marker of tomorrow’s “classic” cars. Level 2 conveys a reasonable mix of human and machine involvement and can save many more lives if mandated in privately owned vehicles. Level 4 is achievable now under self-imposed operational design domain constraints and fits well economically with the mobility as a service model, especially in underserved areas.
In April, U.S. Sen. Richard Blumenthal took a “partly self-driving Tesla road test.” He did so to “highlight federal legislation he says insufficiently protects against potential dangers of self-driving cars.”
The ride was less than satisfactory. A driver from Consumer Reports had to take over in order to avoid a crash at about 65 mph.
Noted autophile / audiophile, Alex Roy, who has been among the most vocal around the dangers of partial automation and associated terminology, reacted immediately:

Since then, Alex penned a piece arguing for a complete abolishment of the SAE levels in favor of two categories: “Human-Assisted Systems” and “Geotonomous” Driving, not dissimilar from the emphases we will propose here. That said, we don’t realistically see the industry moving away from the SAE levels anytime soon, and in any case, taxonomies are typically intended to represent the breadth of possibilities, for better or for worse.
However, as we’ve previously written, when it comes to AV, “a more holistic way of thinking is a requirement, not a choice.” That includes revisiting standards that may have been set years ago and incorporating what we’ve learned since — from both industry insiders and vested outsiders. As a refresh, here are the SAE levels as described on the NHTSA website*:


More Colloquially…
Level 0 is self-evident: no automation. Level 1 is nominal assistance that has been around for a long time through features like cruise control, possibly also the more recent adaptive cruise control. Level 2 expands on cruise control by extending to steering in addition to throttle and braking, however there is a ongoing debate on whether this steering should be for lane keeping versus lane changing. Regardless, in both Level 1 and 2 situations a human is fully expected to monitor the driving environment.
This changes at Level 3, at which point the system is doing the primary monitoring “with the expectation that the human driver will respond appropriately to a request to intervene.” Level 4 and Level 5 require no driver at all with the system handling all driving tasks, but under pre-defined operable conditions. For Level 4 that means constraints like geofences, time of day or weather conditions, whereas Level 5 assumes no limitations whatsoever.
Odd Ones Out
Since the SAE levels were first published four years ago, it has become apparent that not all are necessarily good ideas to strive for in real-world deployment:
Level 1 implies the vehicle is capable of additional safety-focused automation, yet comes up short at the demonstrable cost of savable lives by the technology that is attainable and affordable today.
Level 3 often straddles a dangerous line between machine capabilities and human attention, as further suggested in recent findings about the Elaine Herzberg tragedy in Arizona. While advanced human-machine interaction technology may be able to address the risks of humans straying too far “out of the loop,” it starts to beg the question of why this level of self-driving is even worthwhile — especially if it presents added costs and diminishing safety returns merely for the sake of increased convenience.
Finally, Level 5, as purely defined, is simply unnecessary if not impossible in any meaningful time horizon. Even today, we ground or ban cars, trains and planes in, say, dangerous weather conditions or high density environments.


That Leaves…
Level 0: a dying breed. These are our vintage or classic cars of the future, which is a perfectly acceptable scenario for those of us (including many of us at CARMERA) who still appreciate recreational driving in the right environments, particularly in more manually-driven vehicles that are fun to operate.
Level 2: all of the life-saving, advanced driver assistance features like collision avoidance and automated emergency braking technology, that can be and should be implemented today, potentially under top-down regulation to ensure auto companies have incentives to incorporate them on a level playing field, and don’t feel they have to compete on safety.
Level 4: in our opinion, the most realistic, yet still highly transformational, pursuit in AV. Realistic given the operational design domain constraints (e.g. geofences, weather conditions) and redundancies (e.g. tele-operation, IoT signaling infrastructure) that allow the driver removal with the full reliability of a production, commercial service. Transformational, not just from a safety standpoint, but also from affordability, asset utilization and commuter happiness perspectives. The full — and attainable — promise of autonomous mobility as a service.
Where We Fit In
As a horizontal technology provider in the AV industry, CARMERA is exposed to a lot of different approaches and philosophies. The views above certainly aren’t all-encompassing and will inevitably evolve, but are based in part on what we’ve been seeing and hearing in 2018 from the serious players preparing for commercial products and services, versus just R&D projects or demos. One might even argue that these views conflict with the business interests that a company like ours would have. However, we don’t actually believe that to be true. The vast majority of Level 2 programs, and 100% of Level 4 programs, are depending on high-definition mapping to ensure that both of these autonomy levels are deployed at automotive-grade levels of reliability and passenger experience. We are actively serving both and committed to doing so far into the future.
Regardless, as we say frequently around here, for AV to work it has to work for all. If a more rationalized approach on where to draw the lines for desirable versus undesirable levels of autonomy is another piece of the puzzle to get us there, we are all for it.
Related Reading: Can regulators keep up with the pace of self-driving tech? (Automotive News)