Tesla.com

On August 31, 2026, a Tesla owner asked Elon Musk on X whether Full Self-Driving would ever reliably steer around potholes. Musk replied with two words: "coming soon". Brief as it was, the confirmation mattered — not because "coming soon" from Musk carries a strong track record on this specific feature, but because this time the feature is also named in an official Tesla changelog. What the changelog cannot tell you is whether the neural network running in your car right now can actually do the job consistently. That question has an architectural answer, and the answer points to a software version that hasn't shipped yet.

Pothole Avoidance Has Had Seven Years to Arrive

The paper trail goes back to April 2019, when a Tesla owner asked Musk on Twitter whether FSD could make small steering adjustments within a lane to protect tires from potholes. Musk agreed it was worth building. In February 2020, he responded affirmatively when asked whether Tesla vehicles could create and share "micro-maps" containing road-feature data — including potholes — that the broader fleet could then download and act on. Drive Tesla Canada reported Musk's micro-map yes at the time. Six months later, in August 2020, Musk went further: "Yes! We're labeling bumps and potholes, so the car can slow down or steer around them when safe to do so." A TechTimes 2020 pothole confirmation report documented that statement.

Six years of labeling, scanning, and promising later, the feature still has not shipped as a named, activated capability in production FSD.

What changed in April 2026 is the level of institutional commitment. When Tesla deployed FSD v14.3 through software update 2026.2.9.6 on April 7, 2026, the release notes included, for the first time, an explicit item under the heading "Upcoming Improvements": Not a Tesla App confirmed the v14.3 "Add pothole avoidance" listing. That is a product team making a public, versioned promise — not a CEO's tweet. The problem is that every subsequent build, from v14.3.4 through v14.3.8 (rolling out now as part of the 2026 Summer Update, firmware 2026.26.6.5), still lists the same item in the same section. EVSHIFT confirms pothole avoidance still listed upcoming in the v14.3.8 build. As of September 1, 2026 — nearly five months after pothole avoidance first appeared in the release notes — it has not been broadly activated across the fleet.

Why Detecting a Hole Is Harder Than Detecting an Object

Pothole avoidance sounds trivial. The engineering is not, and understanding why requires looking at what makes potholes specifically difficult for a camera-only autonomous driving system. A PMC study on vision-based pothole detection difficulty outlines the core problem.

Most objects FSD must detect — other vehicles, cyclists, pedestrians, debris — protrude above the road surface. A camera can identify them through brightness contrasts, known shape profiles, and depth estimation from motion parallax. A pothole is the opposite: a depression below the road surface. Its geometry is negative. The camera sees a shadow, a darkened patch, or a texture change — not a raised object — and must infer depth, width, and risk from that much weaker signal. The difficulty compounds under variable lighting (direct sun flattens the shadow that defines a pothole; night driving removes it), at speed (the reaction window shrinks at highway velocities), and on roads whose damage patterns don't appear in training data.

FSD's vision-based architecture — Tesla chose not to use LiDAR or radar in its vehicles — means every depth estimate comes from monocular camera input. Tesla's VP of AI confirmed at the Q2 2026 earnings call that the company achieves safe autonomy with cameras alone, but for negative geometry the signal is weaker and the training data thinner, because potholes are far less uniformly distributed across miles driven than, say, other vehicles.

To activate pothole avoidance as a consistent, fleet-wide feature, FSD must simultaneously complete three tasks in real time: identify the surface depression as a pothole rather than a shadow or a drain cover; estimate its depth, width, and lateral position with enough precision to determine whether steering around it is safe; and execute the avoidance maneuver smoothly enough that passengers don't notice a violent correction. EVSHIFT's coverage of real-time pothole avoidance complexity confirms each task alone is an edge case. All three, at 65 miles per hour (105 km/h), with oncoming traffic, is where six years of labeling have not yet produced a shippable result.

What FSD's Neural Network Is Already Doing Without Being Asked

Here is what makes the current situation interesting: FSD vehicles are already avoiding some potholes. Not because a dedicated pothole avoidance module has been activated, but because the end-to-end neural network has seen enough human driving footage to start generalizing the behavior. Not a Tesla App has documented FSD's emergent avoidance in current builds.

Owners report that FSD builds as far back as v12.3.4 and as recent as v14.2.2.5 occasionally slow down or make subtle lateral adjustments near road hazards. One documented example showed FSD v14.2.2.5 repositioning a vehicle so that a damaged section of road passed between its tires — whether the system classified the hazard as a pothole specifically or simply as an anomaly to avoid is unclear, but the avoidance happened. The underlying mechanism is emergent: the system has absorbed enough footage of human drivers reflexively steering around road damage that it has begun to approximate the behavior without explicit training on potholes as a labeled category.

The catch is consistency. Emergent behavior from a model trained on millions of driving clips is not the same as a reliable, systematically activated capability. FSD v14's end-to-end neural network runs on approximately one billion parameters — the learnable weights that determine how camera input gets translated into driving decisions. At that scale, the model generalizes pothole avoidance in favorable conditions. In challenging conditions — an unmarked road crater at dusk, a half-filled pothole flush with standing water — the behavior drops out. That inconsistency is why the feature remains in "Upcoming Improvements" rather than activated. TeslaHubs covers the v14 vs v15 model scale in detail.

Does Consistent Pothole Avoidance Need FSD v15?

This is the question the five-month changelog gap raises, and it is the question Musk's two-word confirmation does not answer.

FSD v15 is a different architectural generation. Where FSD v14 runs on approximately one billion parameters, v15 is being built around a model roughly ten billion parameters in size — a tenfold increase. Musk confirmed this in an April 2026 post on X, explaining the delay: "Our rate of advancement with the small model has been so fast that the large model has not yet caught up. V15 will be the large model." TeslaHubs reported Musk's April v15 model announcement in full. At the company's Q2 2026 earnings call in July, Tesla VP of AI Software Ashok Elluswamy confirmed that Robotaxi runs FSD v15 early builds. A consumer release is targeted for late 2026 or early 2027.

The parameter count matters for pothole avoidance specifically because of what scaling research suggests about rare-scenario generalization. A model trained on billions of driving miles will see lane changes and traffic lights thousands of times per mile of data. It will see a specific pothole configuration on a specific road surface under specific lighting conditions far less often. A larger model, trained on the same data, generalizes better to low-frequency edge cases — not because it memorizes them, but because its greater representational capacity can extract more nuanced features from ambiguous visual inputs.

Tesla has not announced whether pothole avoidance will ship as a v14 point release or debut in v15. Not a Tesla App, one of the most reliable Tesla software trackers, notes the feature could arrive before FSD v15 — meaning the team believes v14's 1B-parameter model may be sufficient if targeted training and the improved 3D geometry understanding from v14.3's vision encoder upgrade close the gap. That is the optimistic read. The realistic read is that a feature sitting in "Upcoming Improvements" for five months, across multiple point releases, is waiting for something that incremental tuning alone has not yet delivered.

The Prerequisites FSD v14.3 Laid Down

Whatever version ultimately ships pothole avoidance, FSD v14.3 built the prerequisites that make it closer than any prior version. The update that debuted alongside the pothole avoidance changelog listing also rewrote the AI compiler and runtime from the ground up using MLIR — Multi-Level Intermediate Representation, the same compiler infrastructure that powers ML workloads for TensorFlow — delivering a 20% improvement in reaction time across the entire FSD stack. TechTimes covered v14.3's MLIR reaction time gain in detail. Faster reaction time directly expands the window in which an avoidance decision can be executed at speed.

The same update upgraded the neural network's vision encoder specifically to improve 3D geometry understanding in difficult and rare scenarios. Drive Tesla Canada confirmed v14.3's 3D geometry vision upgrade. This is the core improvement for pothole detection: the system's ability to infer depth from monocular camera input — a critical requirement for classifying a shadow as a pothole versus a drain cover versus a discoloration — has materially improved. That improvement carries through every subsequent v14.3.x build, including the v14.3.8 now rolling out in the 2026 Summer Update.

Hardware also plays a role. Not a Tesla App reports HW4 cameras at 4.5x resolution versus HW3, giving the neural network substantially more spatial information per frame with which to resolve pothole geometry. And because pothole avoidance is a pure software capability, no hardware upgrade is required for it to deploy across the FSD-capable fleet — including HW3 vehicles, which began receiving FSD v14 Lite (a distilled version of the v14 architecture) starting in June 2026. TechTimes covered HW3 vehicles receiving FSD v14 Lite in full.

The Fleet Map That Already Exists

Before active steering avoidance, Tesla built a subtler version of fleet-based road intelligence. When Tesla released software update 2022.20 in July 2022, it deployed a feature called Tesla Adaptive Suspension: Model S and Model X vehicles with air suspension began automatically adjusting their ride height before reaching rough road sections, using rough-road map data continuously generated and downloaded from the broader Tesla fleet. Autoevolution reported on Tesla's fleet road-map update when it launched. The release notes stated, as documented by Not a Tesla App in their 2022.20 adaptive suspension notes: "Tesla Adaptive Suspension will now adjust ride height for an upcoming rough road section. This adjustment may occur at various locations, subject to availability, as the vehicle downloads rough road map data generated by Tesla vehicles."

That is the fleet micro-map infrastructure Musk described in February 2020 already operating in production — for suspension adjustment. Extending that same data architecture to active steering maneuvers under FSD is the logical next step and the "longer-term vision" Musk has gestured at for years. Whether Tesla uses real-time camera detection alone, fleet map pre-notification, or a combination of both for the eventual FSD pothole avoidance feature remains unannounced.

If the fleet-map approach is extended to steering decisions, the implications go beyond individual vehicles. A comprehensive, continuously updated map of road-surface conditions generated by millions of daily Tesla trips would constitute one of the most granular real-time road condition databases in existence — one that local governments could, in principle, use to prioritize pothole repair. As Not a Tesla App notes, Tesla has announced no municipal data-sharing plans for this fleet data.

What the Cybercab Changes About This Question

The Cybercab — Tesla's two-seat robotaxi with no steering wheel and no pedals — is scheduled to hold its formal launch event in Austin on September 3, 2026. Teslarati confirmed the Cybercab Austin launch event details. As of August 31, 45 Cybercabs had been registered with the Texas DMV, as TechTimes reported in Texas Cybercab DMV registration coverage.

The absence of a steering wheel changes the stakes for pothole avoidance. In a supervised FSD vehicle, a driver can grab the wheel when the system misjudges a road hazard. In a Cybercab, there is no wheel to grab. FSD's ability to detect and respond to road-surface anomalies — including potholes — becomes a safety question rather than a comfort question when there is no human fallback. Electrek noted this directly on August 31: the vehicle launching this week will be put on public roads perhaps unable to avoid road hazards.

Tesla's Robotaxi fleet currently runs early builds of FSD v15, not the v14.3.x production software available to consumer vehicles. Whether v15's larger model includes activated pothole avoidance for the Robotaxi context has not been announced.

Reading 'Coming Soon' in 2026

Seven years ago, "coming soon" from Musk on pothole avoidance meant the feature was on a list. In 2020, "coming soon" meant engineers were actively labeling road data. In 2026, "coming soon" carries something it didn't in those years: an official product changelog entry, a vision encoder explicitly upgraded for 3D geometry, a compiler rewrite that improved reaction time by 20%, and a 10-billion-parameter model in active development on the Robotaxi fleet.

None of that makes "coming soon" a delivery date. The structural difference between this confirmation and every prior one is that a product team has put pothole avoidance in a versioned changelog under a named FSD version — but the five months that have elapsed since April 7 without that checkbox being ticked suggest that the incremental improvements in v14's 1B-parameter model have not yet crossed the consistency threshold needed to activate the feature fleet-wide.

The most precise answer to "when?" is: in a late v14 point release if v14.3's 3D geometry improvements are sufficient, or in FSD v15 if consistent performance requires the 10-billion-parameter architecture. Musk did not specify which, and Tesla has not announced a date for either. For owners who have spent years noticing that FSD holds the wheel steady through road hazards that any alert human driver would steer around, that ambiguity is familiar — but the underlying engineering work has genuinely never been closer to resolving it.

Frequently Asked Questions

When will Tesla FSD pothole avoidance actually ship?

Tesla has not announced a specific date. Pothole avoidance has been listed under "Upcoming Improvements" in FSD release notes since v14.3 (April 7, 2026) and remains there in the v14.3.8 builds rolling out in September 2026. It could ship in a late v14 point release or debut with FSD v15, which Tesla is targeting for late 2026 or early 2027. The critical question — whether consistent pothole avoidance requires v15's 10-billion-parameter architecture or can be delivered at v14's one-billion-parameter scale — has not been publicly answered by Tesla.

Does Tesla FSD already avoid potholes sometimes?

Yes, inconsistently. FSD's end-to-end neural network has generalized the behavior from human driving footage without being explicitly trained on potholes as a category. Owners have documented FSD builds from v12.3.4 through v14.2.2.5 occasionally slowing down or making lateral adjustments near road hazards. The upcoming feature would formalize and make consistent what the model already does in favorable conditions but not reliably enough to activate fleet-wide.

Why is detecting a pothole harder than detecting another vehicle?

Potholes are depressions below the road surface — negative geometry — rather than objects protruding above it. A camera sees a shadow or texture change rather than a raised shape, making depth estimation significantly harder from monocular (single-camera) input. The detection must also happen fast enough to execute an avoidance maneuver at highway speeds, under variable lighting, and on road surfaces the training data may not have covered. Tesla's FSD v14.3 vision encoder upgrade specifically targets 3D geometry understanding, which is the core prerequisite for this problem.

Will HW3 Tesla owners get pothole avoidance?

Because pothole avoidance is a pure software capability, it requires no hardware upgrade. HW3 vehicles that have received FSD v14 Lite — the distilled version of the v14 neural network architecture that began rolling out in June 2026 — should receive the feature when it ships, alongside HW4 vehicles. HW4's 4.5x higher camera resolution provides a better input signal, but the feature is not expected to be HW4-exclusive.