Cita de: iRobb date=1585092993Genial @lightbug. Qué gran explicación como solución para un problema que muchos al empezar no conseguimos entender de una sola vez.
Agrego que me olvidé, la razón por la que puede o no pasar de largo un obstáculo es por el otro sync que me faltó explicar, comunmente asociado al "sync previo" o el "autoSync". Previo a cada update físico se realiza un sync, esta vez de T -> RB.
Entonces se puede dar algo como esto:
T -> RB (Sync)
FixedUpdate: Se modifica RB
Simulación
Rb -> T (Sync)
Update: Se modifica T --> Render
Update: Se modifica T --> Render
Update: Se modifica T --> Render
Update: Se modifica T --> Render
T -> RB (Sync) <--- Acá se lee que pasan los datos del transform al RB (notar que el RB quedó con los datos del punto 3)
FixedUpdate: Se modifica RB
Simulación ... acá es donde puede estar el problema (dependiendo de las propiedades del collider y de los dos pasos anteriores)
Rb -> T (Sync)
──────────────────────────────────────────────────────────────────────────
Acá dejo un documento interno de los devs de las físicas por si alguno está interesado en la parte del autoSync (está en inglés):
Priorly, each change of a transform resulted in a synchronous update of the corresponding physics object (if present). That always happened on the main thread and could accumulate to
quite a large number over the frame. That approach was easy to maintain, but the downside of always keeping the physics world in sync with the transform hierarchy was that we ended up writing to the physics engine too often, possibly updating some physics objects a lot more frequently than might be needed.
One particular example is moving the root transform of a ragdoll with N rigidbodies in it (=N physics bones). Once you moved the root from script, there was a message sent to update all
the child transforms down in the hierarchy. Each of those child transforms accepted the message, updated itself, and sent out another message to all the children of its own. This
naturally stacked up, leading to O(N 2 ) complexity of updating a ragdoll. In order to solve that, we have introduced a mechanism of deferred physics updates. With this
new approach, we don't write the transform changes immediately. Instead, we store them in a buffer and apply at a later stage. Updates to the same transform are combined together and
applied as a single update to the final pose (e.g. two updates to a transform become a single update right to the final pose in the buffer).
We always sync transforms prior to running the physics simulation in order to make sure we simulate the latest poses. However, you can run the sync process manually too, using the new Physics.SyncTransforms() function. This makes sure that the transforms in Unity and the ones used by the physics engine are exactly the same.
By default, we run in a compatibility mode that would automatically sync before running the operations that required the up-to-date transforms before [Para 2018.3 en adelante esto está desactivado por defecto]. That is:
Before all the physics queries (e.g. Physics.Raycast, etc)
Before reading data back from the physics engine directly (e.g. Rigidbody.position)
Before simulating particles, if they were configured to compute collisions
Before updating the camera flares effect
As mentioned above, synchronising the transforms might be a costly operation so it's recommended to avoid doing it without the need. For this reason, you can even switch off the
automatic sync, via Physics.useAutoSyncTransforns.


pero esta vez se da la transición de B a C, si está puesto next state entonces B cancelará la transción. El resto son variaciones de esto.