Latency Budgets: Why Sensor Speed Matters as Much as Accuracy

A perfectly accurate reading that arrives too late is often worse than a slightly noisy one that arrives on time.

Latency is the delay between something physically happening and the robot's software receiving usable data about it — sensor capture time, data transfer time, and processing time all add up. A balance controller running at hundreds of updates per second has a correspondingly tiny latency budget to work with.

Why this creates real tradeoffs

A more sophisticated perception algorithm (heavier vision models, more thorough sensor fusion) is often more accurate but slower, which can make it a worse choice for a time-critical control loop even though it produces better results in isolation. Engineers frequently trade some accuracy for the guaranteed low latency a safety-critical loop needs.

How it's managed in practice

Many humanoid control architectures split perception across different update rates: a fast, simple loop (IMU and proprioception, updating at very high frequency) handles immediate balance and safety, while a slower, richer loop (full vision and scene understanding) handles higher-level decisions where a few tens of milliseconds of extra delay is tolerable.