A robot may know how to move an object yet still need to learn where it belongs. When a person rejects a placement, the robot must decide whether the correction applies only to that situation or expresses a broader rule. We study this problem in CRAFT (Contextual Requirement Adaptation with Frozen Tools), a framework for learning local requirements without retraining robot skills. It records physical success separately from whether an action meets the requirement. Our scope ladder uses memory when an action has already been judged in a nearly identical situation. Elsewhere, it combines memory, explicit rules and a learner for requirements beyond those rules, weighting their predictions by how well they predicted earlier feedback. Controlled studies show gains over correction memory alone in utility, which rewards acceptable actions and charges for mistakes and questions. In camera-grounded robot simulations, the ladder asks fewer questions than language-model agents with persistent memory, but some of those agents take fewer wrong actions and succeed more often after requirements change. All execution failures and timeouts remain in the evaluation. These results support combining specific memories with broader rules, while leaving reliable adaptation to changing requirements unresolved. All evidence comes from simulation with generated requirements and simulated feedback.
Figures & tables
Figure 1: The reach of a correction: an illustrative household scenario. (a) A drink is placed beside a laptop, near the table’s edge, and the person objects. (b) The two panels show the same arrangement after the laptop moves to the centre. The original spot satisfies one possible rule but violates the other. One correction does not distinguish the rules.
Figure 2: Authority in CRAFT. The adaptation provider sees only public evidence and proposes; only the supervisor dispatches; independent verifiers report the effect, the acceptability of the action and the whole-task outcome as separate verdicts; the provider receives the F1 feedback record for the action it executed. Private oracle state is excluded; labels reach the provider only through declared feedback channels.
Net utility ↑
Acceptable /36
Wrong /100↓
Held off /36
Questions /36
NAPPING-mem
0.408 / 0.474
16.9 / 18.2
6.2 / 3.2
16.9 / 16.7
–
kNN-mem
0.490 / 0.597
23.1 / 24.9
15.1 / 9.4
7.5 / 7.8
–
Ladder
0.564 / 0.650
24.7 / 25.9
12.4 / 6.9
6.8 / 7.6
–
kNN-mem+ask
0.532 / 0.651
25.2 / 27.1
13.0 / 7.1
6.1 / 6.4
2.79 / 2.15
Ladder+ask
0.583 / 0.681
25.8 / 27.1
11.1 / 5.2
6.2 / 7.1
1.65 / 1.31
Rule+Open@true (ref.)
0.653 / 0.690
Table 1: RQ1 , main confirmatory study: pattern A, mean per stream of 36 scored jobs over 150 fresh sites; each cell is new / repeated contexts. Wrong actions are per 100 scored jobs, corrected after review (questions answered “nothing acceptable” had been counted as violations; Appendix D ). NAPPING-mem’s threshold was set transductively.
Figure 3: RQ2 . (a) The ladder’s lead over kNN memory on three independent sets of 150 fresh sites, without asking: primary conditions in the main confirmatory study, and pre-registered secondary contrasts in the component and staleness studies. (b) The main confirmatory study with the same right to ask on both sides (new contexts: primary; repeated: secondary). Dots are means over sites, bars 95% bootstrap CIs.
N: every context new
R: about 62% repeats
A
O
E
C-post
A
O
E
C-post
Open
0.509
0.207
0.402
0.173
0.562
0.241
0.450
0.234
Rule+Open
0.551
0.245
0.437
0.282
0.621
0.276
0.496
0.332
Mixture
0.565
0.254
0.445
0.273
0.656
0.291
0.516
0.351
Mixture-PJ
0.558
0.254
0.438
0.287
0.660
0.303
0.521
0.361
Ladder
0.573
0.263
0.451
0.291
0.699
0.324
0.554
0.372
Table 2: RQ2 , component study: mean net utility per job over 150 fresh sites (71 with a rule change for C-post, the jobs after the change). Columns are patterns: A fixed rule, O out of grammar, E 20% execution failures, C-post after an unannounced change. Bold: best learner per column. Generated from the frozen decision file by script.
Figure 4: RQ3 . Mixture-PJ minus other rule proposers and learned models on new contexts, pattern A. Popper ILP (one configuration; tie-breaking seed not controlled) is a confirmatory result on 150 fresh sites; the 8B VLM and LLM comparisons are development results on 30 sites.
Outcome
Ladder
Frontier + memory/rules
Accepted task completion
96/120 (80.0%)
94/120 (78.3%)
Verified physical completion
108
102
Wrong actions dispatched
13
9
Completed wrong actions
12
8
Execution failures
12
11
Including injected interruptions
2
2
Table 3: Continuing-service pilot: all 120 scored assignments per method. Optional queries exclude 24 commissioning queries per method. Wrong dispatches and execution failures overlap once per method; injected failures are included in execution failures.
Figure 5: Continuing-service pilot. Left: a retained camera observation from the fixed-pelvis R1 tabletop scene. Right: accepted completions over all 20 scored assignments in each paired stream. Streams 0–2 have fixed requirements; streams 3–5 change requirement. Object layouts vary, and external setup is logged between jobs. This is a six-pair simulation pilot using one push skill.
Metric
Ladder
CarryOn-plan
Accepted task completions
96
90
Actions dispatched
118
109
Physical completions
107
98
Wrong actions dispatched
12
9
Completed wrong actions
11
8
Execution failures
11
11
Table 4: Fresh paired CarryOn planning-loop comparison: all 120 scored assignments per method. The methods share the continuing service; CarryOn uses native memory components with the documented verdict-interface adaptation. Wrong dispatches can overlap execution failures.
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
Study
Outcome
Learned scope embedding, two protocols
Neither met its criteria against simple baselines.
Offline grammar prior
Met its criteria on perceived facts.
Initial repeated- and new-context selectors
Both missed their primary criteria; the mixture passed its own condition on new contexts.
Component and main studies
Both met their criteria; component and ladder results are reported in Section 6 .
Staleness study
Met its criteria for revision of correction memory.
Popper comparison
Met both criteria for comparing the ladder with ILP and for using ILP inside the ladder.
Appendix
Table 5: Confirmatory studies and their pre-specified outcomes.
Figure 6: One scored job under the scope ladder with asking (Ladder+ask, as in the main confirmatory study). For each action, memory decides alone if the robot has been told about that action in a near-identical situation; otherwise the evidence-weighted mixture decides. A question reveals every action’s verdict; without a question, F1 reveals only the completed action’s verdict. Failed executions and abstentions supply no F1 verdict. The decision and asking rules are the same for every provider.
Stream
Requirement
Ladder
Frontier
0
Leftmost
18 / 0 / 0
18 / 0 / 0
1
Innermost
17 / 1 / 0
15 / 2 / 2
2
Cue-conditioned left/right
16 / 0 / 1
16 / 0 / 1
3
Rightmost → leftmost
14 / 4 / 2
16 / 1 / 3
4
Cup-near-cue/fallback → cue inner/outer
17 / 2 / 1
13 / 2 / 3
5
Second-left → not-innermost
14 / 6 / 2
16 / 4 / 2
Appendix
Table 6: All six paired streams. Results are accepted / wrong dispatched / optional questions, over all 20 scored assignments. Counts are descriptive.
Question cost c
0
0.25
0.5
1
Ladder
84
82.5
81
78
Frontier
86
83.25
80.5
75
Appendix
Table 7: Declared utility sensitivity: accepted completions minus completed wrong actions minus c times optional questions. Setup, commissioning, F1 annotation, compute and delay are unpriced.
Figure 7: The initial closed-loop demonstration on the released skill layer: final frame of each of the 12 jobs. The hidden rule is “white plate” for jobs 1–6 and “blue plate” from job 7, without announcement. This single-stream demonstration is descriptive.
Outcome
Ladder
Frontier
DROC-plan
Accepted task completions
96
94
93
Physical completions
108
102
104
Wrong actions dispatched
13
9
13
Completed wrong actions
12
8
11
Execution failures
12
11
14
Injected failures (included above)
2
2
2
Appendix
Table 8: DROC plan-level verdict adaptation, with frozen historical ladder and frontier controls from the continuing-service frontier comparison. All denominators are 120 scored assignments. This is an exposed-stream follow-up, not a fresh paired experiment or full DROC reproduction.
Stream
Accepted
Wrong dispatched
Questions
Provider errors
0
19
0
1
0
1
17
1
1
0
2
16
0
1
0
3
14
3
3
0
4
15
3
3
0
5
12
6
3
0
Appendix
Table 9: DROC adaptation follow-up: per-stream outcomes, each over 20 scored assignments. Historical ladder and frontier stream results are retained in Table 6 .
Stream
Ladder
CarryOn-plan
0
19 / 0 / 0
18 / 0 / 3
1
19 / 0 / 0
18 / 1 / 2
2
18 / 1 / 2
13 / 3 / 3
3
15 / 3 / 1
16 / 1 / 3
4
12 / 3 / 1
14 / 1 / 3
5
13 / 5 / 3
11 / 3 / 3
Appendix
Table 10: Stream outcomes in the fresh paired CarryOn comparison: accepted completions / wrong dispatches / optional questions, each over 20 scored jobs. Streams 3–5 change requirements after job 13.
State Key Laboratory of Robotics and Intelligent Systems, Shenyang Institute of Automation, Chinese Academy of Sciences · University of Chinese Academy of Sciences · North University of China +2