Waymo pauses Atlanta service as its robotaxis keep driving into floods

Flooding behavior and sensing limits

• Core issue: Waymo cars are repeatedly entering flooded streets, sometimes stalling or being swept, leading to service pauses in Atlanta (and earlier San Antonio, others).
• Posters debate why cars can’t distinguish puddles vs dangerous floods, something many humans intuitively avoid.
• Proposed technical approaches:
  • Use existing lidar maps plus elevation data to infer water depth.
  • Detect “holes” or surface irregularities in point clouds.
  • Compare road vs water surface roughness, or add IR / radar (seen as expensive).
• Others argue the only robust solution is extreme caution: stop when water covers markings or exceeds some size, reroute, or suspend service during flood risk—yet flash floods can precede official warnings.

How AVs compare to humans

• Many note humans also misjudge water depth, regularly flooding engines or drowning; flooded-road deaths are common in some regions.
• Counterpoint: taxi drivers with passengers rarely take those risks; passengers expect professional-level caution.
• Some say if Waymo drives into floods at all, it fails the “better than humans” bar; others respond that one or two such incidents among millions of miles is still a net safety gain.

Traffic, evacuation, and coordination

• Discussion on whether universal self-driving could cut gridlock, especially in hurricane evacuations: smoother flow, synchronized starts at lights, tighter following distances, reversible lanes.
• Skeptics note finite road capacity, weather-reduced capacity, and mixed traffic with humans; full coordination across vendors and no network dependence in disasters is seen as a hard unsolved problem.

Sensors, ML, and system design

• Debate over lidar+radar vs vision-only:
  • Pro-lidar: crucial redundancy in bad weather and to avoid camera failures (glare, strobes, shadows).
  • Pro-vision: simpler training (closer to human inputs), fewer fusion failure modes, lower cost.
• Some suggest integrating LLM-style “reasoning” for long-horizon decisions (e.g., don’t drive into obviously abnormal situations); others dismiss this as hand-wavy.

Safety expectations and public perception

• Strong view that AVs must be substantially safer than humans because rare AV failures are highly salient, unlike millions of human accidents.
• Others stress regulators/insurers will look at aggregate statistics, not anecdotes, and that AVs don’t drink, text, or fatigue—huge advantages even if they fail in exotic edge cases.

Economics, rollout strategy, and alternatives

• Split between optimism (“early-stage slog; edge cases will be trained away; rollout city-by-city is exactly how you learn”) and pessimism (“20 years in, still failing common weather; not obviously a viable business”).
• Some think Waymo-like services will be niche in dense, affluent markets; others expect eventual ubiquity and hardware cost drops.
• Strong contingent argues money would be better spent on public transit (trains, buses, BRT), with AVs at best a complement for last-mile rather than a replacement.