Waymo robotaxis were left stranded across San Francisco’s Presidio area after the July 4 Golden Gate Bridge fireworks, sitting immobilized through planned road closures until their batteries drained. At least one vehicle caught fire after prolonged idling. The incident exposed a gap between publicly available closure data from multiple government agencies and the routing systems that autonomous vehicles rely on to avoid exactly this kind of trap.
How planned closures boxed in Waymo’s autonomous fleet
The fireworks display on the Golden Gate Bridge was not a surprise. It was part of the Semiquincentennial Fourth of July celebration, and government agencies spent weeks broadcasting the scale of the disruption. Caltrans had already warned drivers that US‑101 would temporarily close in both directions for the event, with significant congestion and planned freeway and connector shutdowns expected across the corridor feeding the Presidio, in a notice posted on its District 4 update. Separately, the Presidio Trust, the U.S. government corporation that manages the park, published detailed access restrictions showing that Halleck Street connecting Crissy Field to the Main Post would close as part of its event guidance. San Francisco’s own event page mirrored those same restrictions and parking limits in its city advisory.
The closures were designed to manage the massive crowd drawn to the bridge for the fireworks and to keep traffic from overwhelming the Presidio’s narrow roads. But the same restrictions that kept human drivers out also locked autonomous vehicles in. Waymo robotaxis that had entered the Presidio before the closures took effect found themselves unable to exit once the access points and connectors shut down. With roads around them barricaded or converted to one‑way flows for pedestrians and shuttles, the vehicles’ normal routing options collapsed.
Without a human driver to improvise, reroute through unfamiliar terrain, negotiate with traffic officers, or simply pull over and wait for a tow, the vehicles sat in place. Waymo’s remote assistance teams, which normally can help vehicles navigate unusual scenarios, had limited options if every algorithmically viable route out of the Presidio was blocked by a hard closure. As the hours passed and the traffic control plan remained in force, the cars idled or powered systems on and off until their batteries eventually died. One caught fire, though no official fire department report or Waymo statement confirming the cause has surfaced in available public records.
The core problem is straightforward. Every piece of closure information that could have prevented this was published days in advance by at least three government entities: Caltrans, the Presidio Trust, and the City of San Francisco. The data existed and was clearly intended to shape how people and vehicles moved in and around the bridge. The question is whether Waymo’s routing and dispatch systems ingested it and, if they did, why they failed to act on it in time.
Government closure data that Waymo’s systems did not act on
The evidence trail starts with the agencies themselves. Caltrans published its transportation notice on June 29, five days before the event, specifying the full US‑101 closure and warning of significant congestion tied to the Golden Gate Bridge fireworks. The Presidio Trust’s planning document listed specific road closures inside the park, including Halleck Street, and provided official guidance for access, parking, and transit. San Francisco’s event page coordinated the same access limits, describing how traffic officers, barricades, and transit reroutes would reshape movement in the northern part of the city. The National Park Service also posted conditions updates for the Golden Gate National Recreation Area, signaling that this was a region‑wide operation rather than an isolated street festival.
All of this information was structured, public, and available well before the first shell launched over the bridge. Caltrans maintains a real‑time traffic tool at QuickMap that shows active closures, lane restrictions, and incidents. The Presidio Trust and San Francisco both published maps and text descriptions that, while aimed at human readers, follow predictable formats that software can parse. For a human driver, checking any one of these sources would have been enough to avoid the area entirely, leave early, or park outside the most constrained zones.
For an autonomous fleet operating dozens or hundreds of vehicles in a dense urban zone, the failure to incorporate this data into routing decisions created a cascading problem. Vehicles entered a closing zone in the late afternoon and evening, accepting ride requests from riders headed to the waterfront and viewpoints. As the traffic control plan ramped up, exit routes began disappearing from the routing graph. Once the core arteries and connector streets closed, vehicles that were already inside the Presidio could not find legal paths out. They then cycled through route‑finding attempts, repositioned short distances, and eventually stopped as their battery reserves ran down. When the fireworks ended and crowds began to disperse, some of the stranded robotaxis had already become immovable obstacles, compounding congestion.
The hypothesis that event‑specific closure data from Caltrans and the Presidio could trigger preemptive repositioning in autonomous vehicle routing systems is not speculative. It describes a basic operational protocol that ride‑hail companies with human drivers already follow informally. Drivers check traffic apps, read local news, and avoid known trouble spots during major events. Dispatchers sometimes geofence areas around stadiums or parade routes to keep drivers from getting trapped.
Autonomous systems should be able to do this faster and more reliably, given that the closure data is published in structured, machine‑readable formats on government websites and feeds. A minimal rule set could have flagged the Presidio as a high‑risk zone during the closure window, stopped accepting trips into the area, and gradually moved idle vehicles to safer staging locations. The fact that Waymo’s fleet did not reposition before the closures took effect suggests either the data was not ingested at all, or the system lacked rules to act on it in a way that prioritized vehicle egress over short‑term ride demand.
This raises broader questions for regulators and the public. If AV companies are not yet reliably integrating planned closure data, then every large‑scale event-marathons, protests, parades, and fireworks-creates a fresh opportunity for fleets to be boxed in. And when those fleets are electric, with finite battery reserves and complex thermal management systems, prolonged idling in constrained conditions can turn from an operational problem into a safety risk.
Gaps in the record and what riders should watch for next
Several important details are still missing from the public record. No primary Waymo statement has addressed how many vehicles were stranded, how long they sat before their batteries died, or what the company’s remote operations team did during the post‑fireworks period. No California DMV incident report has surfaced to confirm the battery depletion timeline or the circumstances of the vehicle fire. Fire department records that might clarify whether the blaze resulted from battery failure, overheating during prolonged idling, or some other cause have not appeared in available documentation.
Real‑time AV telemetry data from the night of the event, which would show exactly when and where vehicles became stuck, has not been made public. No rider or operator statements from any institutional source have described how Waymo’s dispatch handled the situation as roads closed around its fleet. It is not clear whether trips into the Presidio were cut off at a certain time, whether riders were warned that vehicles might be delayed exiting the area, or whether any vehicles were manually recovered before their batteries fully drained.
The absence of these records makes it difficult to determine whether this was primarily a software oversight, a data integration failure, or a deliberate decision to keep vehicles in the area to meet demand despite known risks. Each possibility implies a different remedy. If the systems never saw the closure data, then basic ingestion of public feeds is the first priority. If the data was visible but not acted upon, then fleet‑level policies and safety margins need to be rethought. And if the risks were known but accepted, regulators may ask whether current oversight frameworks adequately account for large planned events.
For riders, the lesson is more immediate. When booking autonomous trips around major civic events, it is worth checking the same closure maps and advisories that human drivers use, and being cautious about trips that end inside tightly controlled zones. Until AV operators demonstrate that they can reliably anticipate and adapt to complex, multi‑agency traffic plans, the safest assumption is that their vehicles are only as informed as the data they actually use-and as conservative as the rules they are programmed to follow.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
