UMTRI-2015-2
Road Safety With Self-Driving Vehicles: General Limitations and Road Sharing With Conventional Vehicles
Why a self-driving vehicle that is safer than a human driver does not automatically produce safer roads.
- Authors
- Michael Sivak · Brandon Schoettle
- Published
- January 2015
- Pages
- 8
- Format
Paper
The argument
The paper takes apart an assumption that was close to universal in 2015: that once an automated vehicle demonstrates a lower crash rate than the average human driver, deploying it necessarily reduces road casualties. The authors point out that this holds only in a fleet composed entirely of such vehicles, and that no plausible transition path produces that fleet for decades. In the interim, automated and human-driven vehicles share the same roads, and the relevant question becomes how the two interact.
Mixed-fleet interaction
Much of the paper is concerned with the failure modes that belong to neither vehicle type alone. Automated vehicles follow rules; human drivers negotiate. An automated vehicle that stops correctly at an ambiguous intersection may be rear-ended by a driver who expected it to proceed. Human drivers extend informal courtesies and read intent from vehicle motion in ways that a rule-following system neither offers nor interprets. The authors treat these as structural features of the transition rather than as bugs to be patched.
Limits the authors set out
The paper is also candid about what automation does not address. It cannot exceed the physics of stopping distance or tyre friction. It does not remove the consequences of poor road design, degraded markings or adverse weather beyond sensor tolerance. And it introduces its own class of failure — the automated system that behaves correctly by its own logic in a situation its designers did not anticipate.
Why it is still cited
A decade of deployment has largely borne the argument out. The operators who reached genuine driverless service did so by bounding their operating domain tightly rather than by solving general driving, which is a practical acknowledgement of exactly the limitations set out here. The paper is cited across academic literature, trade press and vehicle-security research published in eight languages, which is why the PDF is served at its original path rather than reorganised into a tidier URL.
How the paper has been used
The paper's citation pattern is unusual and worth noting, because it explains why the original URL matters. It is cited in the academic traffic-safety literature, as expected. It is also cited heavily in vehicle-security research — including work published in eight languages by a single security vendor — because its analysis of interaction failures maps onto the question of what an attacker could induce a rule-following vehicle to do. And it appears in trade and general press coverage whenever the safer-than-humans claim resurfaces. Three quite different readerships arrived at the same eight pages.
What it got right
Two things, both load-bearing. The first is that the crash-rate comparison is the wrong test during a transition, because the transition is the entire foreseeable future and mixed traffic behaves differently from either fleet alone. The second is that general driving is harder than the comparison implies — and the practical proof is that every operator which reached genuine driverless service did so by bounding its operating domain tightly rather than by matching human performance across all conditions.
What it did not anticipate
The paper reasons about automated vehicles as a growing share of an otherwise normal traffic mix, gradually displacing human drivers everywhere. What happened instead is geographic rather than proportional: within a small number of mapped metropolitan service areas the automated share is now material, and outside them it is effectively zero. That changes the mixed-fleet problem from a national transition into a set of local ones, each with its own road layout, weather and driving culture — arguably a harder problem to generalise about, and certainly a different one.
Citation
Michael Sivak and Brandon Schoettle. Road Safety With Self-Driving Vehicles: General Limitations and Road Sharing With Conventional Vehicles. University of Michigan Transportation Research Institute, report UMTRI-2015-2, January 2015.
The PDF is served at /wp-content/uploads/2015/01/UMTRI-2015-2.pdf, the path it has occupied since publication. Other papers in the same series and the wider public literature are indexed in the report library.
Reference
About this paper
Who wrote UMTRI-2015-2?
Is this the original PDF?
/wp-content/uploads/2015/01/UMTRI-2015-2.pdf, the path it has occupied since publication in January 2015. Those exact addresses are cited in published academic and trade work, so moving them would break citations that have been resolving for a decade.