Tesla FSD 14.3.7 Nearly Hits Train Despite NHTSA Probe

A Near-Miss That Wasn't Supposed to Happen Again

A Tesla running Full Self-Driving software once again approached an active railroad crossing this month and nearly rolled into the path of a passing freight train, according to dashcam footage shared by auto journalist Lei Xing 110. The car was moving at roughly 9 mph with FSD v14.3.7 — Tesla's newest publicly released build — engaged as the crossing's lights flashed and the train's horn blared. Xing says he had to brake manually to keep the vehicle off the tracks 101112. It is a scenario that should be among the easiest for any driving system to recognize: bright warning lights, a lowered or lowering gate, and a massive, loud object bearing down on the crossing. Yet outlets covering the episode, including Futurism and Electrek, note that the car continued toward the tracks anyway, only stopping because a human overrode it 110.

Not a One-Off Incident

The timing made the footage sting more than a typical near-miss clip, because Electrek says it is at least the third time this year it has documented Tesla's FSD approaching a train dangerously 101112. In March, a Model 3 traveling near West Covina, California at about 23 mph plowed straight through lowered crossing barriers positioned roughly at the height of the car's cameras, with the driver braking only around the moment of impact 1314. A month later, in Texas, longtime Tesla owner Joshua Brown says his car — stopped at a crossing with gates down and lights flashing — suddenly accelerated on its own; he floored the accelerator rather than braking, and the car smashed through the gate before clearing the tracks just ahead of the oncoming train, which struck the arm against his window 31415. Brown, who says he has logged more than 40,000 miles on FSD, called it the first time the system had ever let him down 1415. In 2025, a Tesla in self-driving mode reportedly navigated onto tracks in eastern Pennsylvania and was struck by a train after occupants had already exited the vehicle 101416.

NBC News, credited by multiple outlets as the reporter who first surfaced the broader pattern, says it interviewed six Tesla owners about crossing failures and reviewed footage from four of them, while identifying more than 40 additional social-media accounts describing similar problems 101617. One owner, Italo Frigoli, told NBC that when reporters accompanied him back to the same North Texas crossing where his car had previously failed, FSD again failed to react properly to an oncoming train 1617. Reporting on the topic frames these as recurring, reproducible failures rather than isolated flukes, though the tallies rely on self-reported and unverified social posts and should be read as a signal of a pattern rather than a precise failure rate.

Why the Software Version Matters

The latest scare is notable because it happened on Tesla's newest build, v14.3.7, which followed the earlier v14.3 release built on a rewritten MLIR-based compiler that Tesla said delivered 20% faster reaction times, along with explicitly promised improvements in handling "rare and unusual objects extending, hanging, or leaning into the vehicle path" 141819. Commentators, including Electrek, have pointed out that a lowered crossing gate or an approaching train would seem to fall squarely within that description, making the continued failures harder to dismiss as an untouched edge case 101114. Tesla's release notes never explicitly name railroad crossings, so it's not accurate to say the company promised a dedicated fix — but the juxtaposition of updated language and a fresh near-miss has fueled skepticism about how much real-world improvement each version actually represents.

Other recent releases have focused on more visible, less safety-critical changes. FSD v14.1 introduced a much-discussed "Sloth" slower-speed mode and early robotaxi-related features after months of delay, drawing mixed reactions from Tesla-watchers, some of whom flagged sudden-braking complaints during a public livestream demo 568. Other footage has shown FSD 14 completing smooth, low-speed parking maneuvers at a Costco lot in Canada, illustrating the split reality of the software: capable in constrained, low-risk scenarios, but still prone to serious lapses in high-stakes ones 9.

Regulators Are Already Circling

The railroad failures are not happening in a vacuum. NHTSA opened Preliminary Evaluation PE25012 in October 2025 after tying 58 incidents, including 14 crashes, to FSD's behavior — largely centered on the vehicle running red lights or entering opposing lanes of traffic 131720. The agency's own opening summary describes complaints involving vehicles proceeding into intersections against red signals and executing lane changes into oncoming traffic, and notes that some incidents appeared to cluster at the same Maryland intersection, suggesting a repeatable defect pattern 20. The investigation covers an estimated 2.88 million to 2.9 million FSD-equipped vehicles, and by December the tally of documented violations had grown to roughly 80, drawn from dozens of driver complaints, Tesla's own reports, and media accounts 1213.

Senators Ed Markey and Richard Blumenthal have separately pressed NHTSA to examine the railroad-crossing failures specifically, calling them a "disturbing safety risk" 101114. NHTSA has confirmed to NBC News that it is aware of the crossing incidents and has been in contact with Tesla 1617. The agency's information requests to Tesla have sought granular detail — software versions, driver warning timing, and vehicle data starting 30 seconds before each violation — reflecting a focus not just on whether FSD malfunctions, but on whether drivers are given adequate time to intervene when it does 1320.

The Marketing-Versus-Reality Gap

The repeated crossing failures have sharpened a long-running dispute over what Tesla is actually selling. The company markets the system as "Full Self-Driving," yet officially classifies it as an SAE Level 2 driver-assistance feature requiring constant human supervision and readiness to take over 1517. That gap between branding and legal definition is central to a nationwide class-action suit, Waller v. Tesla Inc., which alleges Tesla misrepresented what its FSD hardware and software could ultimately deliver, according to coverage of the litigation 2. Tesla, for its part, publishes safety statistics claiming FSD-supervised driving results in far fewer collisions than the average human driver — as low as one crash per several million miles versus roughly one per 660,000 miles for unassisted drivers, according to the company's own vehicle safety reporting 15. Critics note such aggregate figures say little about the system's performance in specific, dangerous scenarios like railroad crossings, where a low overall crash rate could still coexist with a concentrated, high-consequence failure mode.

A Widening Competitive Field

Tesla's struggles are unfolding as rivals push their own driver-assistance systems forward. Nissan, partnering with UK-based Wayve Technologies, plans to launch an updated version of its ProPilot system by March 2028 explicitly positioned to compete with Tesla's FSD 4. That timeline underscores how much ground competitors believe they can make up even as Tesla's technology remains under federal scrutiny, and it raises the stakes for Tesla to demonstrate that its incremental software updates are closing safety gaps rather than merely adding new features.

The Bottom Line

None of the individual incidents definitively proves that a crash was unavoidable without human intervention, and Tesla has not conceded that its software specifically cannot handle railroad crossings. But the recurrence of near-misses across multiple software versions, vehicle models and states — combined with an active, expanding NHTSA investigation and a pending class-action lawsuit — has made railroad crossings a specific, symbolically loaded test case for Tesla's broader autonomy claims. As Tesla continues rolling out incremental updates and touting new capabilities like robotaxi features and smoother parking, the persistence of a single, well-documented failure mode raises a pointed question that has followed the company through 2025 and 2026: how much has actually changed under the hood when the car still doesn't reliably stop for a train.