One Vehicle Rolled Over at -63°

The vehicle convoy scenario had already been validated — 21 vehicles running in formation without issue. Then we ran it again one day and one vehicle flipped to a roll of −63°. Working yesterday, broken today. That's the most disorienting class of bug.

We Suspected Physics — But

A rollover naturally points to physics — mass distribution? Suspension? Collision geometry? Wheel friction? We spent days staring at vehicle physics parameters. But something was odd: the other 20 vehicles were completely fine, and only one vehicle, only at one specific position, was rolling over.

💡 "Only one vehicle, only at one specific location" was the clue — If the vehicle physics were broken, similar vehicles should fail similarly. A single vehicle rolling over at a specific position is a signal to suspect not the vehicle, but the terrain at that location. The symptom pattern points to where the culprit lives.

The Real Culprit — The Environment Had Changed Silently

Inspecting the terrain revealed the answer. That environment had originally used a validated flat terrain, but a batch job run a few days earlier had silently swapped the environment DB to a different version with steep facets. The vehicle was unchanged; the ground under it had been replaced.

Rendering diagram…

Once we restored the environment DB to the validated flat version, that vehicle's roll came back from −63° to a normal range (4.5°). The physics had been fine all along.

Vehicle roll angle comparison. On the silently-swapped steep terrain, roll reaches 63° (well above the rollover-risk threshold of 45°). After restoring the validated flat terrain, roll is 4.5° — normal. Not a single line of vehicle code was touched.

Lessons — Config Disguises Itself as a Physics Bug

⚠️ "If the environment looks wrong, compare it against the documented canonical" — Some batch jobs don't just fill in missing assets with placeholders; they also replace working assets with the wrong version. And the result looks like a physics bug (rollover, sliding). If an environment seems suspect, before digging into vehicle code, first verify that the environment is on the same version as the validated canonical.

This belongs in the same family as silent fail and speculative gates — a silent change that produces a misleading symptom and sends debugging in the wrong direction. The difference here is that it was not code but data (environment version) that changed silently.

In a nutshell

📌 One vehicle in a validated convoy rolled over at −63°, but the culprit wasn't physics — it was an environment DB that had been silently swapped from the validated flat terrain to a steep version. "Only one vehicle, at one specific location" pointed to the terrain. When an environment looks suspicious, check whether it matches the canonical version before touching vehicle code — config disguises itself as a physics bug.