Laya Warehouse Safety: a robot picks advance, shift or wait above a deterministic shield

# Laya Warehouse Safety

**A typed decision model, a busy warehouse, and one robot trying not to become an incident report.**

</div>

This project is a visual, replayable experiment in using
[Laya](https://github.com/NandhaKishorM/laya) as a fast operational decision layer for an
autonomous warehouse robot. Workers and forklifts move through a deterministic grid. The robot
must reach a loading bay while choosing whether to advance, shift, or wait.

The simulation keeps emergency collision prevention deterministic. Laya chooses operational
actions above that safety layer; it cannot override the safety shield.



## Why this experiment?

Laya evaluates several typed questions over one state in a single model invocation. A warehouse
crossing makes the result visible: a late or poor decision can cause an avoidable stop, detour, or
collision in an unshielded research run.

The included browser replay evaluates these outputs in parallel:

- `action` — `choice`: a normalized distribution over advance, shift left, shift right, and wait.
- `collision_risk` — `score`: low through critical.
- `path_blocked` — `noul`: probability that the direct path is blocked.
- `needs_operator` — `noul`: probability that a person should review the situation.

The controller applies Laya's selected action. Keeping all actions in one normalized choice avoids
comparing probabilities from separately phrased binary questions.

That replay's model was domain-trained only for `action`. It preserves the other three outputs to
demonstrate Laya's typed interface, but they are unvalidated diagnostics. Do not interpret them as
calibrated warehouse risk or escalation signals.

The benchmark harness trains and evaluates only `action` and `path_blocked`. It does not train
`collision_risk` or `needs_operator`,