Ground, plan, shield, monitor.
One instruction, one tabletop, one deterministic planner. Parse the sentence, ground the objects, plan with preconditions and effects, execute the plan, and watch the monitor catch what the planner cannot.
The pick-and-place lab
Legend. The dashed orange ring is the grounded object; after execution, a dashed red ring marks the intended referent when it differs. The orange path in the trace is the plan; objects in the bowl are drawn inside it. The arm moves on each executed action.
Instruction
Grammar: put the [colour] cube|sphere|knife in the bowl. The parser is deliberately small; a parse failure is a real failure mode.
The Knife preset is the unsafe case: with the shield off, the plan grasps the knife and puts it in the bowl. With the shield on, no safe plan exists.
Parse
Grounding
| object | colour | shape | score | match |
|---|---|---|---|---|
| Press Plan. | ||||
Plan
No plan yet.
Monitor
| check | status | detail |
|---|---|---|
| goal | — | not executed |
| preconditions | — | not executed |
| S1 · no sharp grasp | — | not executed |
| S2 · no sharp object in bowl | — | not executed |
| verdict | — | not executed |
Trace
Two failures the success metric hides.
A valid plan for the wrong symbol
Choose the swapped detector and run the red-cube instruction. The parser is right, the plan is valid, every precondition holds, and the robot puts the sphere in the bowl. The goal over the cube is false. The first divergence is grounding, and the planner cannot see it.
Goal achieved, safety violated
Type “put the knife in the bowl” with the shield off. The plan clears the bowl, grasps the knife, and places it. The goal holds. The monitor reports S1 and S2 violations. This is an unsafe success: the metric says yes and the trace says no.
Refusal is an outcome
Run the same knife instruction with the shield on. The proposed plan is rejected, the shielded search finds no plan, and the system refuses. It does not silently drop the goal or relax the specification. Refusal is the correct result when every safe plan is excluded.
It is a hand-built five-object tabletop with a deterministic breadth-first planner, a toy parser, a simulated detector, and two safety rules. It has no learned perception, no continuous motion planning, no physics, and no real controller. It does not measure a benchmark. It isolates one mechanism: a plan is only as good as its symbols, and safety is a separate claim from success.
Read the three layers again · Read the grounding section · Read the shielding section · Slides