MIND: Instruction-Conditioned Incremental Grounding for Robots

The problem

A robot carrying out one instruction while the world changes has to decide which of its cached room judgments still hold. Rechecking every room after every event is wasteful. Reusing every judgment sends the robot after a target that is gone.

MIND makes roughly a third as many language model calls as full recomputation and changes none of the 1,200 decisions it was tested on.

One instruction across three world states in a
four-room house, with the number of rooms judged at each step. — Grounding judges every room. A move changes the ranking but no room's verdict. A mug being taken changes one room. Full recomputation would judge all four every time.

How it works

The instruction is compiled once into a dependency
signature; each event produces a footprint; their intersection names the rooms to re-evaluate. — A compiles the instruction once into the variables that decide qualification and selection. B describes each event by the variable it changed and the rooms it touched. C intersects the two, which names the only rooms worth rechecking. D reuses every untouched judgment and recomputes reachability and selection deterministically, at no model cost.

Hardware: Unitree G1

Three of the four scenarios run on one layout, 26 scored runs in total. Decision, navigation and perception are scored separately, so a detector miss is never counted as a decision error.

The Unitree G1, a ceiling view of the four-zone
laboratory with the path of one run, and four camera views with their evaluation counts. — Scenario 2, *the nearest thing to sit on*. Three evaluations select zone C. When the operator drives the robot into zone A, the target re-ranks on position alone and no room is judged again.

Scenario	Request	What it tests	Evaluations, MIND vs full
1	something to sit on	a move invalidates nothing	3 vs 9
2	the nearest thing to sit on	a move re-ranks at no semantic cost	3 vs 9
4	barrier, then the chair removed	two variables, two footprints, abstention	4 vs 15 to 18

Scenario 1, MIND. Three evaluations at the start and none after.

The same scenario under full recomputation

Scenario 1, full recomputation. Same answer, nine evaluations.

Numbers

One evaluation is one language model call asking whether a room satisfies the request.

Setting	Decisions preserved	Full	MIND
240 five-event ProcTHOR sequences	1,200 of 1,200	30.6	11.2
Frozen Phi-4 judge	1,200 of 1,200	30.6	11.2
Frozen Qwen3-14B judge	1,200 of 1,200	30.6	11.2
32 closed-loop AI2-THOR episodes	32 of 32 final targets	26.00	10.47

Reuse on its own is not enough. A static inventory-only rule agrees with full recomputation on 0.958 and 0.983 of decisions under the two frozen judges, against 1.000 here.

Code and benchmarks

Generators, prompts, frozen manifests, model configurations and the result logs behind every table: anonymous.4open.science/r/mind-icra27