(Image: iStock)
I confess. I wasn’t aware of Waymo’s superhuman driver model —NIEON – until Waymo recently brought it back to life after one of its robotaxis hit a child in a school zone in Santa Monica. Waymo cited NIEON as evidence to argue the “material safety benefit of the Waymo Driver.”
Waymo posits that the accident was unavoidable, because even a superhuman driver model like NIEON would have hit the kid.
But wait. What exactly is NIEON?
As it turns out, Waymo unveiled NIEON (Non-Impaired, with Eyes always ON) in 2022, asserting that it represents “an ideal human state for driving,” according to the company.
From an engineering perspective, NIEON is a useful concept. Engineers feel more comfortable having an established “model” against which they can benchmark their own technology.
For marketing and communications, NIEON offers convenient fodder to exonerate robotaxis that go haywire and crash.
In the Santa Monica case, fortunately a minor incident, Waymo argued that “the Waymo Driver outperformed the NIEON human driver model (always attentive and doesn’t get distracted or fatigued), by avoiding more collisions and mitigating serious injury risk in simulated fatal crash scenarios.”
Say what?
The event occurred when the pedestrian suddenly entered the roadway from behind a tall SUV, moving directly into our vehicle’s path. Our technology immediately detected the individual as soon as they began to emerge from behind the stopped vehicle. The Waymo Driver braked hard, reducing speed from approximately 17 mph to under 6 mph before contact was made.
To put this in perspective, our peer-reviewed model [NIEON] shows that a fully attentive human driver in this same situation would have made contact with the pedestrian at approximately 14 mph. This significant reduction in impact speed and severity is a demonstration of the material safety benefit of the Waymo Driver.
The emphasis in bold above is mine.
(Source: Waymo)
Peer-reviewed research
Introducing NIEON in 2022, Waymo blogged:
“We are releasing two new scientific, peer-reviewed papers that present methods to compare autonomous vehicle performance to human driving—an important component of determining the readiness of autonomous driving systems.”
The“peer-reviewed” papers cited include a The Response Time paper that presents the “framework Waymo uses for analyzing and modeling response timing in a crash-imminent situation on the road,” and a Collision Avoidance Benchmarking paper showing “a novel methodology to evaluate how well autonomous driving systems avoid crashes.”
Given that NIEON has been available to expert scrutiny for five years, I was curious about the identity of its peer-reviewers, and what sort of traction the superhuman driver model is getting from other companies who want to benchmark their AV’s road-readiness.
Asked about NIEON, Missy Cummings, professor at George Mason University, told me, “I have never seen this model peer-reviewed beyond very narrow scenarios (see https://www.sciencedirect.com/science/article/pii/S0001457524000058).”
She added, “I know of no one other than Waymo to use it.”
Phil Koopman, safety expert and emeritus professor at Carnegie Mellon University, was not among the peers who reviewed NIEON, either.
However, Koopman did mention NIEON in his book on Embodied AI Safety published last year.
In Section 7.4.2, Koopman warns that a “behavioral requirement based solely on sensor data detection of obstacles would miss critical contextual information.”
Models like NIEON, in his view, “are useful, but do not on their own address the full scope of avoiding negligent driving behavior related to a failure to properly react to high-risk context.”
What’s wrong with NIEON?
Maybe it’s just me. But I couldn’t quite swallow Waymo’s claim that even a superhuman driver, like NIEON, would have hit the kid anyway—and harder. Such a conclusion, drawn from a proprietary simulation and glossed with “scientific” verbiage, strikes the ear as dubiously convenient.
How do we know that NIEON genuinely models the best possible behavior of a human driver?
Asked about Waymo’s NIEON-inspired logic, Koopman said: “I think the argument boils down to their computer reacting faster than a typical human driver in Perception-Reaction Time.” Koopman noted that this calculation “does not take into account context and defensive driving.”
Koopman isn’t alone. Many experts question NIEON’s awareness of context.
MIT research scientist Bryan Reimer, acknowledging that the NIEON model provides a strong reference point for a ‘superhuman’ driver, stressed that it “does not fully capture contextual human judgment, such as heightened caution in school zones, or the way passengers and environmental cues can support safe driving.”
Cummings suspects that Waymo’s push for NIEON is part of a larger effort to deflect blame. “If Waymo wants a useful comparison,” she suggested, “they would ask the school or local police for statistics on how many kids have been struck by cars around that school in that location.”
Bottom ine
We can discuss the pros and cons of theoretical superhuman drivers forever, arguing endlessly over whether the superhuman model would have beat the Waymo Driver to the brake pedal. But to get to the bottom of the Santa Monica case, said Cummings, Waymo “should absolutely release the video” of the event.
She added, “The Minneapolis ICE interactions have shown how different one organization’s interpretation of an event differs from the actual video. Waymo claims it is transparent. If so, it should release the video so we can all see for ourselves.”
Related story:
The flaw? Kryptonite!