2.4 · Estimation, Time, and Partial Observability
Physical decisions are made from delayed, noisy, multirate evidence. State estimation and time synchronization are therefore core components of the learning system, not optional preprocessing.
- Describe prediction and correction in state estimation.
- Quantify how timestamp error becomes spatial error.
- Diagnose missing history, missing sensing, and model underuse.
Prediction and correction
A state estimator propagates a prior through a motion model and corrects it using new observations. The resulting belief combines evidence and uncertainty, giving the policy a better information state than a raw snapshot.
Time is a physical coordinate
At nonzero velocity, timestamp error becomes position error. Multirate sensors require a reference clock, hardware timestamps when possible, bounded interpolation, and a documented policy for stale measurements.
Partial observability as a design choice
A larger network helps when relevant information exists in the current observation or history. It cannot recover an arbitrary hidden contact mode with no informative evidence. Add sensing, active perception, conservative action, or an explicit belief.
When information is absent or stale, model scale is not a substitute for observability.
