Waymo updates 3,800 robotaxis after they 'drive into standing water'
Waymo has issued a software-based safety recall for 3,800 robotaxis after some vehicles drove into standing water, highlighting how difficult it is for autonomous systems to distinguish harmless puddles from dangerous flooding. Commenters debate technical approaches — from high-definition lidar maps and water sensors to pure vision models — and note how fast-changing road conditions, weather, and rare edge cases strain current mapping and perception systems. The incident also reignites broader questions about how “recalls” should be defined for over‑the‑air updates and whether self‑driving fleets can realistically outperform human drivers across both everyday and extreme scenarios.
Standing Water as a Hard Perception Problem
- Many note that distinguishing wet pavement, shallow puddles, and deep or moving water is genuinely hard, even for humans.
- Others dispute that humans “frequently” drive into floodwaters, saying it’s rare relative to total miles driven and more tied to specific regions and conditions.
- Examples from Texas, England, rural US, and fords/low-water crossings underline how common dangerous water situations can be in some areas.
Mapping, HD Maps, and a Dynamic World
- Waymo is described as heavily reliant on HD lidar-based maps of service areas.
- Some argue this mapping could help infer water depth by comparing current readings to stored road geometry.
- Others stress maps go stale quickly due to construction, lane shifts, floods, sinkholes, earthquakes, etc., and real-time aggregation and distribution of changes is unsolved at scale.
Sensors and Technical Approaches
- Suggestions: dedicated water / wading sensors, ultrasound, float switches, moisture sensors, radar, world-model-based inference, or inferring from vehicle deceleration and crowdsourced phone sensor data.
- Concerns: many sensors only detect depth after entering water; condensation, salt, bumps, and interference complicate designs.
- Lidar often treats standing water like a mirror; multi-return lidar may sometimes see both surface and road.
- Debate over lidar+camera vs camera-only: more sensors can help but also dilute engineering focus; fusion and sensitivity/specificity trade-offs are nontrivial.
Safety, Edge Cases, and AV vs Human Drivers
- Some see this as an expected edge case; software can be patched fleetwide, leading to long-term safety improvements humans don’t get.
- Others highlight regressions and novel edge cases will always exist, and “stopped car” failure modes can still be dangerous (e.g., in floods, on highways, on tracks).
- Comparisons to Tesla: anecdotes show FSD sometimes avoids water and sometimes aims for it, implying remaining model/data limitations.
“Recall” Terminology and Regulation
- Multiple comments note that “recall” here means a safety defect plus a fix, often just an over-the-air update.
- Several argue the term misleads the public, conflating catastrophic hardware defects with software patches, but others emphasize it is a regulated, legally defined term that signals a safety issue regardless of fix method.