Rigidbody uses both interpolation and extrapolation to give an impression of smooth motion while simulating at a comparatively low frequency. Internally, this is implemented by calculating the transform poses every update. In the case of interpolation, the last two simulated poses are used to calculate a new transform pose for this frame. In the case of extrapolation, the last simulated pose and velocity are used instead. Since it is designed to be lightweight, however, we don’t communicate these poses back to the physics engine. The poses are only presented to the systems outside of physics (e.g., graphics and animation). Because of that, for instance, a raycast won’t detect a body at the interpolated pose. 

To keep physics from noticing the transform changes, the mechanism was to have a Physics.SyncTransforms() call each update right before pose write, followed by an internal method call to clear all transform updates for physics. That caused two classes of problems:

1. If a scene has at least one interpolated body, all the transform changes to all the physics components were synced with the physics engine each update (even though they’re mostly needed once per FixedUpdate); and
2. If a change was made to a transform that had a Rigidbody component with interpolation on it, interpolation for this object broke because the user-made transform change was propagated to the physics engine and effectively changed the last simulated pose (the pose is not stored separately – it’s just the pose that the physics engine uses currently).

To address these problems, we updated the interpolation code so that it doesn’t need to sync all transforms for each frame. This change also improves performance; the new interpolation code runs faster than before (depending on the scene complexity).