Decorum in AI and Robotics

• Decorum is the alignment of actions with context-specific norms, integrating traditional rules with subjective aesthetics and ambient cues.
• It is applied across AI tasks such as table setting, robot speech, and 3D scene synthesis to capture both conventional and personalized preferences.
• Evaluation spans objective metrics like RMSD, FID, and subjective ratings, balancing technical performance with human-centered judgments.

Decorum, as treated in recent AI and robotics research, denotes the alignment of action, arrangement, or expression with context-specific norms of appropriateness. In a table-setting task, this includes both “traditional rules of Western table setting decorum” and personal judgments about “the color, shape, or material of each object”; in robot speech, it appears as behavior rated as “socially appropriate for the scene” and “aware of the surrounding ambiance”; in indoor scene synthesis, it appears as language-conditioned control over room style and layout (Newman et al., 2024, Tuttosi et al., 2022, Marshall et al., 23 Mar 2025). Taken together, these works treat decorum less as a fixed etiquette manual than as a composite of conventional regularities, ambient and social constraints, and user-specific preferences.

1. Conceptual foundations

In household rearrangement, decorum is introduced through the claim that many tasks “can only be considered fully successful when they account for personal preferences such as the visual aesthetic of the scene.” The canonical contrast is between a solution optimized according to “traditional rules of Western table setting decorum” and a solution that also accounts for object-specific visual properties. The former is rule-based and conventional; the latter incorporates personal aesthetic preference, including “color, shape, or material of each object” (Newman et al., 2024).

This framing places decorum at the intersection of normative and subjective structure. The same source characterizes such preferences as subjective, fine-grained, and continuous, and notes that they can be influenced by “cultural norms or decorum.” The implication is that decorum-sensitive behavior cannot be reduced to task completion alone. A table may be correctly set in a conventional sense while still failing to satisfy the intended visual aesthetic.

In human–robot interaction, an analogous notion appears in vocal adaptation. There, decorum is operationalized through judgments such as “This person’s voice is socially appropriate for the scene,” “This person knows how they should present themselves in this ambient context,” and “This person is aware of the surrounding ambiance.” The relevant variables are not table geometry or object identity, but ambient and social context: quiet versus noisy, formal versus casual, calm versus exciting (Tuttosi et al., 2022).

A plausible implication is that, across embodied AI, decorum is increasingly modeled as latent appropriateness inferred from context, examples, and priors, rather than as a small set of hand-written rules. This is especially visible where language is described as ambiguous or cumbersome for specifying fine-grained preferences.

2. Personal and conventional decorum in multi-object arrangement

The most explicit formalization appears in preference-conditioned table setting. A preference model $\mathcal{M}$ is defined over an initial table state $a_0$, a set of candidate objects $O$, a language prompt $\ell$, and a history of prior examples $K$, with the objective

$\max_{a\in A} \mathcal{M}(a~|~K, O, a_0, \ell).$

An arrangement of length $T$ is represented as

$$a_{1:T} = [(\mathbf{x}_1, \mathbf{o}_1), \dots, (\mathbf{x}_T, \mathbf{o}_T)],$$

where $\mathbf{x}_t \in \mathbb{R}^3$ encodes continuous placement and rotation $(x,y,r)$, and $a_0$0 encodes discrete object attributes such as color, material, and type. Preference history is represented implicitly as prior arrangements and placement orders rather than as an explicit utility function or fixed decorum rules (Newman et al., 2024).

The same formulation distinguishes a human preference-generating process $a_0$1 from the learned model $a_0$2: $a_0$3 Generic decorum is not encoded as hard constraints. Instead, it is folded into pretrained vision-and-LLM priors and into the task prompt. When the prompt says “set the table,” the model can draw on prior knowledge of typical table settings, while the in-context examples $a_0$4 condition the model toward user-specific deviations.

Technically, this system uses zero-shot prompting with GPT-4o, Claude-3 Haiku, and Gemini 1.5 Pro. It introduces lifting functions $a_0$5 and $a_0$6 that convert arrangements and candidate objects into images and text descriptions so that the vision-LLM can reason visually. A central device is to convert continuous placement into a multiple-choice spatial problem by overlaying a labeled grid on the table image and asking which cells an object will occupy and which cardinal direction it will point. Four variants are defined by combining object and arrangement representations: LOUMA, LOGMA, MOUMA, and MOGMA. DegustaBot-MOGMA with GPT‑4o is reported as the best-performing configuration for the main naturalistic experiments (Newman et al., 2024).

The dataset contains 995 table arrangements from 199 participants, each completing 5 trials. Each trial uses one of five different empty tables and offers 125 objects, with 25 each of plates, cups, forks, knives, and spoons. Participants can choose specific instances, drag-and-drop objects anywhere, rotate them arbitrarily, and use at least 3 objects. Success is measured by object prediction accuracy, RMSD after best-fit alignment with the Kabsch algorithm, and subjective acceptability. The subjective evaluation calibrates RMSD against self-ratings of perturbed scenes, using a normalized rating threshold of 0.2 to define acceptability. Reported mappings include 72.7% acceptable for RMSD $a_0$7, 33.3% for RMSD $a_0$8, and 19.9% for RMSD $a_0$9. On real user data, RMSD improves from 0.174 with $O$0 to 0.116 with 2 examples and 0.108 with 4 examples; object prediction accuracy rises from 3.7% to 9.0% and 8.9%. The paper summarizes the resulting preference-alignment level with the statement that “50% of our model’s predictions are likely to be found acceptable by at least 20% of people” (Newman et al., 2024).

These results support a distinction between coarse and fine decorum. The system often captures symmetry, relative positioning, and place-setting structure, but it less reliably matches fine-grained stylistic choices such as exact object instance, color, or material. This suggests that conventional spatial regularities are easier to transfer than individualized aesthetic taste.

3. Social and ambient decorum in robot speech

In robot speech, decorum is treated as the ability to “read the room” and adapt vocal style to the surrounding ambiance. The study uses food-service scenarios and six recreated ambiances: fine dining, café, lively restaurant, quiet bar, noisy bar, and nightclub. In Zoom-based data collection, these ambiances are defined by sound, visual background, textual scenario, and social tone. In the on-robot study, they are recreated with projected images, lighting, and sound, with ambient levels of 50 dB for fine dining, café, and quiet bar, 55 dB for lively restaurant and noisy bar, and 60 dB for nightclub (Tuttosi et al., 2022).

The methodology has three stages: collecting and validating human voice data in Zoom ambiances, clustering vocal utterances to identify primary voice styles, and testing robot voice styles in recreated ambiances. The human dataset contains 837 utterances, reduced to 685 after validation and filtering. Validation retains only passages for which the majority of raters assign at least 4 on all questions loading onto social appropriateness and ambiance awareness. Feature extraction yields 10 audio features spanning loudness, spectral properties, and rate of speech. K-means clustering is then used to identify three interpretable voice styles: a calm cluster associated with fine dining and quiet bar, an exciting or bright cluster associated with lively restaurant, and a loud Lombard-like cluster associated with nightclub (Tuttosi et al., 2022).

These clusters are mapped to Pepper’s TTS voices. The calm cluster is mapped to Neutral, the exciting or bright cluster to Joyful, and the loud cluster to a modified Didactic voice in which pitch is increased by 130% to approximate a Lombard-style configuration. Measured Pepper voice characteristics are reported as follows: Neutral has pitch 350 Hz, pitch range 575 Hz, and speech rate 4.18 syll/s; Joyful has pitch 408 Hz, pitch range 699 Hz, and speech rate 4.41 syll/s; Didactic has pitch 367 Hz, pitch range 558 Hz, and speech rate 3.71 syll/s; the modified Lombard voice has pitch 435 Hz, pitch range 467 Hz, and speech rate 3.71 syll/s (Tuttosi et al., 2022).

The main user study includes $O$1 participants. After each interaction, ratings are collected on 7-point Likert scales for social appropriateness, ambiance awareness, comfort, human-likeness, and competency. The choice of voice style significantly affects perceived decorum and related judgments. In fine dining, Neutral is the clear first-choice voice, with counts 60 for Neutral, 33 for Joyful, and 27 for Lombard. In nightclub, Joyful is most often first-choice, with counts 58 for Joyful, 26 for Neutral, and 36 for Lombard. One-way ANOVAs show that Neutral is significantly more socially appropriate than Joyful and Lombard in fine dining, whereas Joyful is significantly more socially appropriate in lively restaurant, noisy bar, and nightclub. For example, social appropriateness means are 5.30 for Neutral, 4.30 for Joyful, and 3.97 for Lombard in fine dining, and 3.88 for Neutral, 5.46 for Joyful, and 3.28 for Lombard in nightclub. Correlation analyses further show strong positive correlation between social appropriateness and ambiance awareness in all ambiances, and substantial positive associations with competency, comfort, and human-likeness (Tuttosi et al., 2022).

The resulting notion of decorum is explicitly relational. A neutral, restrained voice is decorous in a formal and quiet setting, but the same voice becomes decorously inadequate in a nightclub or lively restaurant. Vocal decorum is therefore not a stable property of a voice preset; it is the fit between prosodic profile and ambient-social context.

4. Stylistic decorum in language-conditioned 3D scene synthesis

In indoor 3D scene generation, decorum appears as style-conditioned control over both layout and object appearance. The system called Decorum addresses the problem of generating a plausible, fully furnished 3D room from a room footprint and a textual description, with explicit attention to visual features and style preferences. The core design principle is language-first representation at every stage. Geometry is represented via CSS-style layout code, style via textual descriptions, and objects via tags plus text (Marshall et al., 23 Mar 2025).

The pipeline has three components. Prompt2Ann maps a free-form user prompt and room dimensions to a dense scene annotation with object tags and detailed scene- and object-level text. Ann2Layout maps the annotation to a CSS-style layout that specifies positions, sizes, and rotations. DecoRate performs object retrieval by selecting textured meshes from 3D-FUTURE based on the same annotation. Prompt2Ann and Ann2Layout are built on LLaMA 3.1 8B with LoRA fine-tuning, while DecoRate uses LLaVA-NeXT for multimodal likelihood-based retrieval (Marshall et al., 23 Mar 2025).

The object retrieval formulation is explicitly probabilistic. Given a text description $O$2, the selected object is

$O$3

equivalently

$O$4

with scoring function

$O$5

Here $O$6 is reported as effective. The purpose is to retrieve furniture that matches detailed style descriptions rather than only coarse category-level similarity (Marshall et al., 23 Mar 2025).

The datasets are 3D-FRONT, 3D-FUTURE, and 3D-GRAND. Following LayoutGPT, the system restricts experiments to rectangular rooms. After filtering, bedroom splits are 3397 / 453 / 423 for train/val/test, and living room splits are 690 / 98 / 53. Evaluation uses FID, CLIP-FID, KID, and a multimodal-LMM-based Textual Fidelity Rate (TFR). Bedroom results report FID 18.0, CLIP-FID 1.7, and KID 0.005 for Decorum, compared with 21.0 / 2.4 / 0.007 for InstructScene and 22.5 / 2.5 / 0.006 for CLIP-Layout. For object retrieval, DecoRate achieves Top-10 accuracy 47.3%, Top-5 accuracy 35.0%, and Top-1 accuracy 16.3%, compared with 15.3%, 8.4%, and 0.5% for CLIP. Bedroom mean TFR is reported as 66.6% for Decorum and 50.0% for CLIP-Layout; living-room mean TFR is 72.2% for Decorum and 50.0% for CLIP-Layout (Marshall et al., 23 Mar 2025).

Here decorum is stylistic coherence rather than etiquette in the narrow sense. A room described as “cozy Scandinavian” or “industrial loft-style” is treated as correct only when the generated layout, object set, and retrieved meshes jointly instantiate that style. This expands decorum from social rules toward scene-level consistency between language, geometry, and appearance.

5. Representation and evaluation across domains

Across these studies, decorum is never measured by a single scalar native to all tasks. Instead, each domain defines a representation for appropriateness and a task-specific set of evaluation procedures.

Domain	Representation of decorum	Evaluation
Table setting	Preference history $O$7, object set $O$8, prompt $O$9, arrangement $\ell$0	Object Prediction Accuracy, Geometry RMSD, subjective acceptability
Robot speech	10 audio features, three clustered voice styles, recreated ambiances	7-point ratings for social appropriateness, ambiance awareness, comfort, human-likeness, competency
3D indoor scenes	Prompt2Ann, Ann2Layout, DecoRate, CSS layout, tagged annotations	FID, CLIP-FID, KID, TFR, Top-$\ell$1 retrieval accuracy

In table setting, evaluation combines exact object identity with geometric structure and user-rated tolerance for perturbation. The RMSD metric is minimized after best rigid alignment with the Kabsch algorithm, which emphasizes relative layout rather than absolute coordinates. This is consistent with the paper’s interpretation of decorum as arrangement structure rather than pointwise placement (Newman et al., 2024).

In robot speech, decorum is evaluated perceptually. The critical questions concern whether the voice is socially appropriate, whether Pepper is aware of the ambiance, and whether the interaction induces comfort, competency, or human-likeness. The empirical finding that social appropriateness and ambiance awareness are strongly positively correlated in all ambiances indicates that decorum is treated as jointly social and situational rather than purely acoustic (Tuttosi et al., 2022).

In 3D scene synthesis, decorum is evaluated by distributional realism and text-scene fidelity. The proposed TFR explicitly compares the likelihood of the intended description under the generated scene against the same description under negative scenes: $\ell$2 This makes decorum measurable as relative textual alignment rather than only image realism or object-category correctness (Marshall et al., 23 Mar 2025).

A plausible implication is that decorum-sensitive AI requires mixed evaluation regimes. Conventional geometric or perceptual metrics are insufficient on their own when success depends on personal preference, ambient fit, or stylistic coherence.

6. Limits, variation, and homonymous technical uses

The present literature emphasizes that decorum is highly variable and only partially captured by current models. In table setting, preferences are described as “extremely high-dimensional and idiosyncratic,” and the participant pool is located in the United States, with no explicit modeling of cross-cultural etiquette. The work is limited to top-down simulated table surfaces and does not evaluate other household domains such as shelves, desks, or living rooms (Newman et al., 2024).

In robot speech, the voice corpus is based on 8 female undergraduate actors in Zoom ambiances, and the on-robot study uses Pepper with a limited TTS control interface. The authors explicitly note limits involving virtual versus real ambiances, group dynamics, robot constraints, speaker demographics, and static rather than real-time voice selection (Tuttosi et al., 2022).

In indoor 3D scenes, Decorum is restricted to rectangular rooms and to single bedrooms and living rooms. The paper also notes dependence on the style coverage of 3D-FRONT, 3D-FUTURE, and 3D-GRAND, as well as runtime costs for full-inventory multimodal retrieval (Marshall et al., 23 Mar 2025).

The same orthographic root also appears in unrelated technical systems. DECOR in room acoustics denotes “Deep Exponential Completion Of Room impulse responses,” a model for synthesizing late reverberation from the first $\ell$3 of an RIR (Lin et al., 2024). DecoR in large-language-model routing denotes “Decomposition-based Routing,” a capability-aware framework that matches new queries against historical logs to balance predictive performance and computational cost (Lv et al., 25 May 2026). Decor in decentralized learning denotes a differentially private variant of decentralized SGD that uses pairwise-canceling correlated Gaussian noise under SecLDP (Allouah et al., 2024). These usages do not address etiquette or aesthetics directly. This suggests that, in current arXiv usage, “decorum” functions both as a classical notion of appropriateness and as an attractive acronymic label for systems organized around structured conditioning, decomposition, or completion.