In January, Waymo claimed its robotaxi hit a child at 6 mph, while a computer model suggested an attentive human driver would have struck the child at 14 mph, TechCrunch reported. The January incident highlights Waymo's strategy: simulating human driving to prove its autonomous systems are safer. Yet, human decision-making is complex and variable; any model remains an imperfect, potentially biased, benchmark. Companies like Waymo are now defining AV safety metrics through these human-like benchmarks, a move likely to accelerate public acceptance and regulatory approval, but one that may also obscure the full spectrum of real-world human driving nuances.
How Waymo's 'Reference Driver' Simulates Human Reactions
- The new Reference Driver (ReD) model reconstructs human behavior in pre-collision scenarios, even simulating driver 'surprise', Zamin Uz reports.
- ReD also incorporates a 0.2-second pause to simulate human single-foot operation of gas and brake pedals, according to The Verge.
These details reveal Waymo's meticulous approach to modeling human physical limitations and reaction times. However, by focusing on such granular delays and specific cognitive responses, the model risks overlooking the adaptive cognitive strengths human drivers often exhibit in novel, complex situations.
Benchmarking Against Humans: A New Standard for Safety Claims
Waymo's new ReD model refines its strategy of using simulated comparisons to assert AV safety. This sophisticated tool reinforces the company's narrative of superior safety over human drivers, building on past claims like the January child-incident scenario. Waymo's consistent reliance on internal models, from that initial claim to the ReD model, marks a strategic pivot: simulated comparative performance is now the primary driver for public trust, eclipsing reliance on real-world incident data.
The Academic Partnership Behind the Model
Scientists at TU Delft, collaborating with Waymo, developed this new computational model to predict human driver responses in hazardous traffic, TU Delft confirms. This partnership lends crucial scientific credibility to Waymo's modeling of complex human driving, moving beyond mere proprietary simulations. It provides a robust, scientifically-backed framework for evaluating AV performance against human benchmarks.
Open-Sourcing the Future of AV Safety Research
Waymo is releasing the Reference Driver model's research code under an academic, non-commercial license, TechCrunch reports. This move could standardize how AV safety is evaluated, fostering broader research into human-like driving models. However, by open-sourcing ReD, Waymo strategically attempts to co-opt the academic community into validating a benchmark that, by design, may inherently favor its autonomous systems, blurring the lines between independent research and corporate advocacy.
If Waymo's human-like benchmarks gain widespread academic and regulatory acceptance, the industry will likely accelerate public trust in AVs, even as the nuances of real-world human driving remain a complex, evolving target.
