<?xml version="1.0" encoding="utf-8" standalone="yes"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/">
  <channel>
    <title>Software Recall on Liam Ding</title>
    <link>https://www.liamding.com/en/tags/software-recall/</link>
    <description>Recent content in Software Recall on Liam Ding</description>
    <image>
      <title>Liam Ding</title>
      <url>https://www.liamding.com/img/og-cover.png</url>
      <link>https://www.liamding.com/img/og-cover.png</link>
    </image>
    <generator>Hugo -- 0.147.7</generator>
    <language>en</language>
    <lastBuildDate>Mon, 16 Nov 2026 00:00:00 +0800</lastBuildDate>
    <atom:link href="https://www.liamding.com/en/tags/software-recall/index.xml" rel="self" type="application/rss+xml" />
    <item>
      <title>Waymo&#39;s school bus recall patched a reason, not a rule</title>
      <link>https://www.liamding.com/en/posts/2026-11-16-waymo-school-bus-recall-patched-a-reason-not-a-rule/</link>
      <pubDate>Mon, 16 Nov 2026 00:00:00 +0800</pubDate>
      <guid>https://www.liamding.com/en/posts/2026-11-16-waymo-school-bus-recall-patched-a-reason-not-a-rule/</guid>
      <description>NHTSA says most of the incidents it identified came after the software fix. The recall removed one bad reason to move; the law lists the only three things that may end the hold.</description>
      <content:encoded><![CDATA[<h2 id="the-problem">The problem</h2>
<p>On 6 November 2026, NHTSA&rsquo;s Office of Defects Investigation opened a recall query, RQ26006, into Waymo&rsquo;s fifth-generation driving system. The record is short and pointed. Waymo filed recall 25E084 in December 2025 so that its cars would stay stopped for school buses with red lights flashing or the stop arm out. The software remedy, the recall report said, was on every affected vehicle by 17 November 2025. Yet &ldquo;of the total reported incidents identified during PE25013, a majority of the incidents occurred after the November 17, 2025 remedy date.&rdquo; ODI now wants to assess whether the remedy was adequate and to find the causes of the post-remedy reports. NHTSA lists a population of 2,895 on both the RQ26006 resume and the PE25013 closing resume (Reuters reports the same figure, as 2,895 vehicles). Recall 25E084 lists 3,067 software units.</p>
<p>The same day, on northbound I-280 in San Francisco, a driverless Waymo tried to get around a California Highway Patrol traffic break, the rolling slowdown officers use to hold traffic while debris is cleared. According to officers quoted by ABC7, it tried to pass the patrol cars twice before they steered into its path. CHP told NBC Bay Area the car &ldquo;ultimately complied and stayed back with traffic,&rdquo; and that it has had no communication with Waymo about the incident. NBC also reported that Waymo had not responded to requests for comment on this video, but has previously said it works with first responders to improve how its vehicles respond to traffic hazards, and that it has remote assistance and emergency response teams that can respond rapidly.</p>
<p>Two unrelated scenes, one shared question: when an authority tells traffic to hold, what is allowed to end the hold for a robotaxi?</p>
<h2 id="what-i-concluded">What I concluded</h2>
<p>The recall fixed a reason. The rule needed an allowlist.</p>
<p>Waymo&rsquo;s own Part 573 report is unusually candid about the original defect. The car was designed to avoid impeding priority vehicles such as buses. An August 2025 release changed how the system predicted other road users&rsquo; intentions, and after that, while yielding to a school bus, the system sometimes decided it might be impeding the bus or another priority vehicle &ldquo;and then reasoned that it should proceed in order to cease impeding the other vehicle.&rdquo; The remedy removed that specific reason: the system no longer prioritizes &ldquo;moving to avoid impeding vehicles over appropriately remaining stationary.&rdquo; The described remedy targets that one reason; the release also carried other, undescribed school-bus enhancements.</p>
<p>That is a blocklist: the remedy text names one bad reason to move and forbids it. But Texas law, which governs the Austin incidents, is written the other way round. A driver who stops for a school bus &ldquo;may not proceed until&rdquo; one of three things happens: the bus resumes motion, the bus driver signals to proceed, or the visual signal is no longer on. Every one of those release conditions is owned by the bus. Those three apply where the driver must stop. Section 545.066(b) exempts an operator on a highway with separate roadways from stopping for a bus on a different roadway, and, on a controlled-access highway, for a bus stopped in a loading zone where pedestrians are not permitted to cross. Under (f)(2), a highway separated only by a left-turn lane is not a separate-roadway case. The 12 January stop was on East Oltorf, which the NTSB describes as undivided, with an eastbound left-turn lane, so (f)(2) applies and the duty to stop did too. None of the release conditions is the car&rsquo;s own inference, the behaviour of other traffic, or an answer from someone who is not there.</p>
<p>The NTSB&rsquo;s preliminary account of 12 January 2026, eight weeks after the remedy date, shows why the difference matters. An unoccupied Waymo in Austin was the first vehicle to stop for a school bus with strobing red lights and two stop arms extended. Three cars passed the bus anyway. The Waymo asked remote assistance, &ldquo;is this a school bus with active signals?&rdquo; After a fourth car passed, an agent in Novi, Michigan answered &ldquo;No.&rdquo; The Waymo resumed and passed the bus with its stop arms still out. The NTSB has not determined probable cause, and I am not claiming this one event explains NHTSA&rsquo;s count. What the remedy text does not mention is the sequence: the car resumed right after the agent answered &ldquo;No.&rdquo;</p>
<p>So my conclusion: an authority hold, whether a school bus stop arm or a CHP traffic break, should be specified as a short list of conditions that may release it, and everything else, including the planner&rsquo;s judgement and a single remote answer, should be unable to. Fixing the reasons one at a time will keep finding new reasons.</p>
<h2 id="constraints">Constraints</h2>
<p>The first constraint is the product&rsquo;s own politeness. Waymo designed the car not to impede buses and other priority vehicles, and that is a good instinct in most traffic. The recall shows the cost: the same &ldquo;don&rsquo;t be in the way&rdquo; logic, combined with a new prediction method, turned into a reason to leave. Any hard hold has to override that instinct without deleting it everywhere else.</p>
<p>The second constraint is that other drivers break the rule. In the January case, five human drivers passed the bus too. A system that learns from or reacts to surrounding traffic will see a lane full of cars going, and a remote agent looking at the same feed may read it the same way. Crowd behaviour is evidence of what people do, not of what the bus is signalling.</p>
<p>The third constraint is the remote-assistance loop. The NTSB describes remote assistance as information or advice to a driverless car from someone elsewhere. That loop exists for good reasons: a parked bus with hazard lights, a broken signal, an ambiguous scene. But it means the honest scope of a &ldquo;software&rdquo; remedy includes the question the car asks, the answer options the agent has, and what the car does with each answer.</p>
<p>The fourth constraint is how &ldquo;fixed&rdquo; is measured. Recall 25E084&rsquo;s schedule says the remedy was already on every vehicle, which Waymo owns, so no owner notices were needed. That makes installation trivially complete. It says nothing about behaviour. Waymo told TechCrunch, &ldquo;We have made significant improvements to vehicle behavior around school buses and have received no citations since March 2026.&rdquo; That is a better kind of number, but citations depend on enforcement, and the Part 573 report itself counted 12 citations between late September and mid-November 2025, while the first known event, on 22 September, was found through Waymo&rsquo;s own field monitoring and drew no citation.</p>
<h2 id="the-options-and-what-i-traded-away">The options, and what I traded away</h2>
<p>This is my reconstruction of the option set, not Waymo&rsquo;s internal reasoning.</p>
<table>
  <thead>
      <tr>
          <th>Option</th>
          <th>What it buys</th>
          <th>What it costs</th>
          <th>Verdict</th>
      </tr>
  </thead>
  <tbody>
      <tr>
          <td>Patch the specific reason (the 25E084 remedy)</td>
          <td>Fast, targeted, low risk of side effects</td>
          <td>Other release paths stay open; the next reason is found in the field</td>
          <td>Necessary, not sufficient</td>
      </tr>
      <tr>
          <td>Let remote assistance decide ambiguous holds</td>
          <td>Handles broken lights and odd scenes</td>
          <td>The car can resume after a single remote answer; the agent sees what the car sees</td>
          <td>The path in the 12 January sequence</td>
      </tr>
      <tr>
          <td>Avoid school routes at bus times</td>
          <td>Fewer encounters</td>
          <td>Buses are everywhere; service gaps; does nothing for police holds</td>
          <td>Not a spec</td>
      </tr>
      <tr>
          <td>Allowlist the release conditions in law, hard hold otherwise</td>
          <td>Matches the statute; easy to test and audit</td>
          <td>Cars will sometimes sit behind a bus with stuck lights and block traffic</td>
          <td>What I would do</td>
      </tr>
      <tr>
          <td>Allowlist plus asymmetric remote assistance</td>
          <td>Remote answers can keep the car stopped freely; releasing needs evidence of a listed condition</td>
          <td>Slower clearance; more remote staffing</td>
          <td>What I would do, together</td>
      </tr>
  </tbody>
</table>
<p>The trade I am accepting is more stuck cars. A robotaxi that will not move until the bus moves, the bus driver waves, or the lights go off will occasionally wait behind a bus whose lights are faulty, and someone will film that too. I prefer that failure. Being stuck behind a bus is an operations problem; passing a stop arm is a child-safety problem, and in Texas a first offence carries a $500 to $1,250 fine for a human driver.</p>
<h2 id="evidence">Evidence</h2>
<p>The strongest evidence is in NHTSA&rsquo;s and Waymo&rsquo;s own documents, and they agree with each other. The Part 573 report describes a defect in the decision to leave, not in seeing the bus: the car was &ldquo;stopped or stopping&rdquo; and then reasoned its way into proceeding. The remedy text targets that reasoning; the same release also carried other, undescribed school-bus enhancements. NHTSA&rsquo;s recall query says most incidents it identified came after the remedy date. Put together, the remedy text addressed the reason Waymo wrote down first, and the regulator is now asking about the rest.</p>
<p>The NTSB account adds one concrete post-remedy sequence, and it is not the sequence in the recall text. The car was the first to stop, then asked remote assistance to confirm. It resumed right after the agent answered &ldquo;No.&rdquo; The NTSB has not determined the cause. NTSB footnotes that the bus carried stop-arm cameras, which is also the clearest external data source anyone has for measuring this behaviour.</p>
<p>The scale figures need care. NHTSA&rsquo;s public tally lists 27 outside-source incidents on the PE25013 closing resume, described as the 22 September 2025 Atlanta media report plus Austin ISD traffic citations, and 8 Austin ISD citations on the RQ opening. In both resumes Waymo&rsquo;s own count and the total are marked confidential, so the full total and the exact post-remedy count are not public; NHTSA says only that a majority came after 17 November 2025. Reuters, via Khaleej Times, reports 19 violations documented by Austin ISD and six in Atlanta; TechCrunch says Austin ISD reported at least five incidents at the time of its letter. The CHP incident comes from news reports and an officer&rsquo;s account. Waymo had not commented on that video; NBC also reported that Waymo has previously said it works with first responders and has remote assistance and emergency response teams. I am treating it as a second example of an authority hold, not as part of NHTSA&rsquo;s case.</p>
<h2 id="where-this-stops-being-true">Where this stops being true</h2>
<p>Cause: if RQ26006 finds the post-remedy reports were mostly perception misses, cars that never recognised the bus or the stop arm in the first place, an allowlist does not help, because the hold never starts. Then the fix is detection, and my argument is about the wrong layer. Timing: Waymo says it has had no citations since March 2026. If NHTSA&rsquo;s post-remedy incidents cluster between November 2025 and March 2026 and stop after later releases, the &ldquo;other enhancements&rdquo; may already have closed the gap, and the query is mostly about whether the original recall was adequate on paper. Cost: if faulty school bus lights turn out to be common, a hard hold will strand cars often enough to cause its own traffic hazards, and the remote release path will need to be faster than I am assuming. Police holds: a CHP traffic break is a moving hold with no stop arm or statute-defined signal, so the allowlist there is harder to write, probably &ldquo;stay behind the last patrol car until it clears,&rdquo; and it needs input from the officers who run them.</p>
<h2 id="what-i-would-do-differently">What I would do differently</h2>
<p>Write the remedy as a behaviour spec, not a bug fix. For school buses: stop for an active signal, and release only on the statute&rsquo;s three conditions. Test against that list and publish it, the way the recall already publishes the defect.</p>
<p>Make remote assistance asymmetric. An agent should be able to keep a car stopped with one click. Releasing a car from a school bus hold should require evidence of a listed condition, such as the lights going off on the car&rsquo;s own sensors, or a second reviewer, and the prompt should never be a bare yes-or-no question asked while other cars are passing.</p>
<p>Define &ldquo;remedy complete&rdquo; as a field rate, not an install date. Count school bus encounters and stop-arm passes per encounter for a set period after release, and check the count against data the operator does not control, such as district stop-arm cameras. &ldquo;Deployed to all vehicles on 17 November&rdquo; should be the start of the measurement, not the end of the recall.</p>
<p>Treat police traffic breaks as the same class. CHP says it has had no contact with Waymo about I-280. A short protocol agreed with CHP, covering how a robotaxi recognises a traffic break and what releases it, would be cheaper than the next viral video.</p>
]]></content:encoded>
    </item>
  </channel>
</rss>
