Snapshot Interpolation

Source: https://gafferongames.com/post/snapshot_interpolation/ Author: Glenn Fiedler | Date: 2014-11-30 | Series: Networked Physics

When to use

If simulation is not deterministic or player counts grow past lockstep comfort, send snapshots of visual state. Receiver does not run the simulation; reconstructs a visual approximation.

Snapshots

Example CubeState: interacting, position, orientation. Include 16-bit sequence; discard older-than-latest on receive. Naive render-latest still hitches because Internet jitter clusters packets.

Lower send rate (e.g. 10 pps) cuts bandwidth but looks choppy without buffering.

Interpolation buffer

Hold snapshots briefly so statistically you have both current and next sample. Trade added latency for smoothness. Linear interp of position shows 1st-order discontinuities; katamari attachments can "pulse" through the parent cube.

Hermite interpolation

Send linear velocity; Hermite spline matches positions and velocities at endpoints → smoother. Orientation: slerp usually enough; angular velocity often unnecessary for cubes.

Packet loss

Snapshots are time-critical but not requiring reliability. Never TCP. Skip lost snapshots; interpolate toward more recent ones in buffer. Rule of thumb: buffer delay ≈ 3× send interval for 2–5% loss (survive two losses) + small jitter margin. At 10 pps → ~300ms + ~50ms = ~350ms. Extrapolation can cut delay but fails badly for nonlinear rigid-body collision (floor tunneling, spring mispredict, katamari tangent vs rotate).

Reduce delay

Increase send rate: 30 pps → ~150ms buffer; 60 pps → ~85ms. Requires bandwidth optimizations (next article).