Custom sensors

The sensor overview covers the user-facing side: attach a sensor with scene.add_sensor(), step the simulation, and read a tensor. This section is for the other side of that boundary. Read it when the built-in sensors are not enough and you need to add your own type, or when you want to understand what happens between scene.step() and the value sensor.read() hands back.

The two pages differ in audience. The overview treats a sensor as a device you configure and query; here a sensor is a small pipeline you implement. The distinction matters because the pipeline is where imperfections, delay, history, and idempotent reads come from: behavior you rely on as a user but must reproduce correctly as an author.

• Sensor pipeline: the runtime model. It follows one measurement from the physics state through the per-step pipeline to read(), and explains why read() is a constant-time memory lookup rather than an acquisition.

• Writing a custom sensor: the author’s guide. It covers which hooks to override, the shape and dtype contracts each must honor, and the automatic plugin registration that pairs an options class with its sensor class.

Start with Sensor pipeline if you want the mental model first; jump to Writing a custom sensor if you already understand the pipeline and want the contract.