FAQ
Is neural perception the default?
No. On the default path object evidence comes from the classical wrist camera pipeline
(ColourDepthBackend: colour + depth), not from a learned model, and not from simulator ground
truth — ground truth is demoted to /perception/ground_truth/* for the error evaluators, and the
overhead camera contributes obstacles only (its object pipeline stays off). Learned models are
future sidecars behind the same observation contracts.
Does the reasoner improve success rate?
Not necessarily. It is off by default, advisory only, and measured mainly for safe degradation / latency, not as a reliability booster.
Why did my goal go to another robot?
Interactive launch-* recipes do not lease a ROS domain. just demo and benchmarks do.
If another baseline is running on the same domain, goals can land on the wrong composition.
Where is the internal engineering documentation?
Kept privately, outside this repository. This Docusaurus site (website/, just docs /
just docs-serve) is the public-facing manual.