A newly published study in the Journal of Machine Design reports that the appearance of a humanoid robot can be systematically evaluated and improved for user acceptance by combining Kansei engineering, questionnaire-based quantitative analysis, and Stable Diffusion model training. The research, conducted by Deng Jun, Li Xiaoyu and Yuan Jiajun at the School of Urban Design, Wuhan University, focuses on humanoid robot designs used in daily life and commercial service settings. It seeks to address a persistent challenge in humanoid robot development: many current designs are shaped primarily by engineering requirements and can appear cold, rigid or distant to users. The study argues that a lack of affinity in humanoid robot appearance may trigger resistance and reduce willingness to use such systems, making affinity a practical design target rather than a purely aesthetic concern.
The research positions affinity as a measurable relationship between product features and users’ emotional responses. For a humanoid robot, affinity is expressed through exterior characteristics such as form, material, surface treatment and color. The study therefore translates users’ affective impressions into design requirements and evaluation scores, then uses those scores to guide the selection and production of training images for a Stable Diffusion model. The reported outcome is a workflow in which a humanoid robot design team can generate a large number of affinity-oriented appearance concepts in a short time, compare them through a scoring mechanism, and identify high-affinity options more efficiently.
- Why Affinity Matters for Humanoid Robot Acceptance
The paper begins from the observation that humanoid robot applications are expanding across service and daily-life scenarios, and that more frequent collaboration between people and robots is becoming likely. In such interactions, appearance influences both emotional response and functional expectation. A humanoid robot that appears approachable may encourage trust, while a humanoid robot that appears harsh or unfamiliar may create psychological distance. The study notes that humanoid robot design has often been dominated by an engineering perspective, producing overall appearances that can be perceived as stiff and lacking in affinity. This is not a minor stylistic issue: the research links low-affinity humanoid robot appearance to user resistance and lower use intention.
Affinity design, as described in the study, emphasizes alignment with users’ physiological and psychological factors and aims to create pleasant, comfortable and relaxed emotional experiences. For a humanoid robot, affinity is not treated as a vague impression but as a set of externalized product features. The research draws on Kansei engineering, a representative approach for exploring relationships between products and emotions, because its goal is to convert affective cognition into engineering design elements. In this framework, affinity can be expressed through a humanoid robot’s modeling, material and color, and those features can be evaluated, scored and used as design guidance.
Previous work cited by the study already points toward several affinity-related features for a humanoid robot. Coordinated overall proportions, smooth contours, simple modeling and a cute style can support affinity. From a color perspective, gentle tones can create an approachable and bright visual atmosphere. From a material perspective, soft materials can produce impressions of affinity and safety. The new study builds on these observations but goes further by quantifying the relationship between humanoid robot appearance features and affinity dimensions, and by using that quantification to train artificial intelligence models for design generation.
- Study Framework Links Humanoid Robot Appearance to Affinity
The research adopts a design-oriented workflow. It uses morphological analysis to extract design elements such as shape, material, texture and color. It then uses the semantic differential method to measure the relationship between affinity perception and those design elements. Questionnaire surveys and quantitative analysis produce affinity feature scores for different design elements. Those scores form a precise design requirement and evaluation mechanism. The next stage uses the score table to guide the screening and production of Stable Diffusion training samples. After training with suitable parameters and multiple iterations, the study generates several model versions and selects the best-performing model. The selected model can be used alone or combined with other style models to generate a matrix of humanoid robot appearance schemes, which are then evaluated with the affinity score table.
The study’s questionnaire design began with 30 pairs of image-related words associated with affinity. These pairs were grouped into three categories: social impression, temperament and style, and vitality. Social impression words describe the closeness between people and a humanoid robot. Temperament and style words describe the robot’s style and character. Vitality words describe the robot’s liveliness and activity. A professional group discussed the pairs and removed those that were semantically similar, obscure or difficult to use when evaluating humanoid robot design elements. Twenty-five design professionals then voted and selected three pairs judged most suitable for describing humanoid robot affinity: “indifferent–affinity,” “cold-hard–gentle,” and “serious–lively.” These correspond to three evaluation dimensions used throughout the study: affinity, gentleness and liveliness. Affinity directly measures how close the humanoid robot feels to users; gentleness reflects care and consideration from the perspective of temperament and style; liveliness reflects vividness and activity from the perspective of vitality.
The study states that questionnaires were used to collect user feedback across these three dimensions, and the analysis examines differences and associations among them. This matters for humanoid robot design because a design feature may increase one dimension while reducing another. The study does not treat affinity as a single undifferentiated impression. Instead, it shows how affinity, gentleness and liveliness relate to each other across different appearance elements, providing more specific guidance for humanoid robot design decisions.
- Design Elements and Affinity Measurement
The questionnaire materials abstracted humanoid robot appearance so that the influence of individual design elements could be examined precisely. For head shape, the study summarized five common forms: square, circular, vertical racetrack, horizontal racetrack and semicircular. For eyes, the study used four forms: circular, square, vertical racetrack and horizontal racetrack. For body proportion, the study referred to infant, child, juvenile and adult proportions and drew planar forms, which were grouped into slim and sturdy types. For shell material, the study used plastic, fabric, metal and transparent material, with two surface effects for each material. For color, the study selected three cool hues, three warm hues and neutral colors including white, gray and black, with saturation and brightness adjusted for each hue. These materials allowed the research to measure how specific humanoid robot appearance features relate to affinity.
The study collected 645 valid questionnaires. Statistical methods were selected according to data type and research purpose. The reported influence of design elements on affinity, from high to low, was head shape, facial expression, color, material and body proportion. The research used Kendall correlation analysis for ordinal variables including head shape, eye shape, body proportion, material and surface treatment, and color across the three affinity dimensions. It used Pearson correlation analysis for continuous variables such as material type and warm-cool color saturation. The results showed strong consistency among the three dimensions for head shape, eye shape, body proportion, material type and color: affinity, gentleness and liveliness were strongly positively correlated. The study concludes that the three measurement dimensions are sufficient to describe affinity accurately. It also reports that surface treatment and the saturation contrast of warm colors showed a negative correlation between affinity and liveliness, meaning that affinity can increase as liveliness decreases in those specific comparisons.
| Design element | Affinity–gentleness correlation | Affinity–liveliness correlation | Gentleness–liveliness correlation |
|---|---|---|---|
| Head shape | 0.356 | 0.303 | 0.284 |
| Eye shape | 0.318 | 0.345 | 0.368 |
| Body proportion | 0.160 | 0.113 | 0.152 |
| White, gray and black | 0.294 | 0.319 | 0.301 |
The study used Kruskal-Wallis H nonparametric tests for multi-group comparisons of head shape, eye shape and body proportion across the three dimensions. Values greater than 3 were treated as positive evaluations. For head shape, the horizontal racetrack form was reported as the most affinity-oriented, followed by the circular form; other forms were not reported as affinity-oriented. For eye shape, the vertical racetrack form was the most affinity-oriented, followed by circular and square forms; the horizontal racetrack eye shape was not affinity-oriented. For body proportion, affinity from high to low was reported as infant proportion, child slim proportion, juvenile slim proportion, child sturdy proportion, juvenile sturdy proportion, and adult sturdy proportion. The study also used Mann-Whitney U nonparametric tests for two-group comparisons of material and surface treatment. Fabric was reported as the most affinity-oriented material, with plush texture described as higher than woven textile texture. Plastic and transparent materials were comparable in affinity, and matte surfaces were higher than glossy surfaces. Metal material, although relatively lively, was not reported as affinity-oriented. The study notes that increasing surface roughness can improve affinity by reducing the cold impression of smooth surfaces.
For body height and color, the study used Spearman correlation analysis. The results showed that greater height was associated with lower affinity. The negative correlation was stronger for sturdy bodies than for slim bodies. For neutral colors, white had the highest affinity, followed by gray, with black the lowest. For warm colors, a position farther to the left in the reported color-space representation was associated with greater affinity; for cool colors, a position farther to the right was associated with greater affinity. Single-factor linear regression was used to compare warm and cool colors across dimensions. The reported order of affinity from high to low was low-saturation warm color, white, high-saturation warm color, low-saturation cool color and high-saturation cool color. The study concludes that higher brightness and lower saturation can increase humanoid robot affinity, and that warm tones such as beige and light orange can create a warm and comfortable impression. Increasing brightness and reducing saturation can also improve the affinity of cool colors.
- Key Findings for Humanoid Robot Appearance
The study’s summary of humanoid robot appearance affinity is organized around modeling, material and color. In modeling, simple, smooth and rounded forms tend to produce亲切 and gentle visual impressions; the English report describes these as approachable and gentle visual impressions. A horizontal racetrack-shaped face, vertical racetrack-shaped eyes and a slim juvenile body proportion were reported to have a more significant positive influence on affinity. In material, soft and textured materials can produce stronger affinity. Surface treatment that increases texture and roughness and reduces the coldness of smooth surfaces can improve affinity. The study recommends combining multiple affinity-oriented materials to enrich the visual effect and make the humanoid robot appear more vivid. In color, higher brightness and lower saturation can produce higher humanoid robot affinity. Warm colors such as beige and light orange can create warmth and comfort, while increasing brightness and reducing saturation can make cool colors more approachable.
To quantify the affinity level of training samples, the study produced an affinity score table for design element features. The table assigns scores to feature codes and provides a basis for screening and producing training samples. The reported score table includes scores such as A1 = 3, A2 = 3, A3 = 0, A4 = 3, A5 = 0; B1 = 1, B2 = 1, B3 = 3, B4 = 0; C1 = 3, C2 = 2, C3 = 4, C4 = 3, C5 = 2, C6 = 1, C7 = 0; D1 = 1, D2 = 2, D3 = 3, D4 = 3, D5 = 0, D6 = 0, D7 = 1, D8 = 1; and E1 = 4, E2 = 2, E3 = 2, E4 = 1, E5 = 3, E6 = 0, E7 = 0. These scores are used as a design evaluation mechanism and as guidance for humanoid robot training image selection.
| Design element area | Reported affinity-oriented feature | Reported lower-affinity or non-affinity feature |
|---|---|---|
| Head shape | Horizontal racetrack form; circular form | Vertical racetrack form; semicircular form |
| Eye shape | Vertical racetrack form; circular and square forms | Horizontal racetrack form |
| Body proportion | Infant proportion; child slim proportion; juvenile slim proportion | Adult sturdy proportion |
| Material | Fabric; plush texture; matte surface | Metal material for affinity |
| Color | Low-saturation warm color; white; high-brightness and low-saturation tones | High-saturation cool color; black in neutral range |
- From Affinity Scores to Stable Diffusion Training Samples
The study then converts the affinity findings into a Stable Diffusion training process. Stable Diffusion is described as an artificial intelligence image-generation technology based on latent diffusion models. It includes a text encoder such as Clip, a latent diffusion model containing Unet, and an autoencoder decoder such as Vae. Its principle is to repeatedly denoise an average noise map through a model function until a clear image is produced. The study notes that Stable Diffusion alone, relying only on prompts, struggles to generate humanoid robot designs with strong affinity and cannot satisfy detailed or specific drawing requirements. This limitation motivates model training based on the affinity score table.
The training samples determine the image-generation effect. When selecting samples, the study ensured that samples matched affinity features and had similar styles but different details, so that the trained model could generalize. To ensure training quality, the study used Stable Diffusion modules including ControlNet, image-to-image and inpainting, combined with Photoshop, to optimize initial materials from overall composition to detail. Through repeated cycles of model training, image generation and modification, the materials gradually met the affinity feature requirements. The study also describes manual annotation after automatic annotation. To ensure accuracy in affinity design feature annotation, the samples were annotated again according to overall modeling, detail modeling, modeling style, material and surface treatment, and color. One reported annotation example was: “affinity robot, minimalism, organic form, science fiction, humanoid robot, oval head, vertical eyes, yellow eyes, no mouth, white body, plastic material, shiny joints, teenage figure, full body, standing, arms at sides.”
Training proceeded through multiple iterations. In each round, samples were modified according to generation results, and each round’s training samples were scored with reference to the affinity score table to improve the affinity of the next result. The first round of training samples and results matched affinity-oriented humanoid robot head and facial features but were too cartoon-like and lacked realism. The second round added more realistic samples, but the generated body proportions became too short. The third round modified the humanoid robot body proportion to match the highest-affinity slim juvenile proportion and added more design details. After training, the generated results basically met the design requirements. The representative training sample scores were reported as 12 in the first round, 14 in the second round and 16 in the third round.
| Training round | Head shape code | Eye shape code | Body proportion code | Material code | Overall color code | Affinity score |
|---|---|---|---|---|---|---|
| 1 | A4 | B2 | C3 | D1 | E1 | 12 |
| 2 | A4 | B3 | C1 | D2 | E5 | 14 |
| 3 | A4 | B3 | C3 | D2 | E1 | 16 |
- Training the Humanoid Robot Model
The study used Dreambooth to train the main model and generate different types of design schemes. It trained multiple Lora models as style models to assist in generating schemes with different characteristics, balancing model training quality and efficiency. Dreambooth is described as a method that adjusts the weights of all layers in the Stable Diffusion neural network, learning features and styles from sample images to meet design needs. Lora is described as a method that uses a small number of images to train a model by inserting a new computation layer into Unet to influence image generation. Compared with Dreambooth, Lora models are smaller and faster to train. The study’s training process includes training sample production, training sample annotation, training iteration, and model screening and use.
The training parameters reported by the study include Learning Rate, Iteration, Batch Size, Epoch, Optimizer, Scheduler, DIM and Alpha. The parameter settings were Learning Rate 1 × 10−4, Iteration 10, Batch Size 5, Epoch 10, Optimizer 8bit-Adam, Scheduler Cosine, DIM 128 and Alpha 64. The study also reports Loss values as an important numerical indicator for evaluating training effects. Models are output after each Epoch, and the dynamic change in Loss values is recorded. The Loss value first decreases and then increases. When the Loss value decreased to around 0.08 at the seventh to ninth Epoch, the study treated this as an ideal reference value. Training parameters and sample size can be adjusted to improve the Loss curve and bring it closer to the ideal state.
| Parameter | Reported setting | Parameter | Reported setting |
|---|---|---|---|
| Learning Rate | 1 × 10−4 | Optimizer | 8bit-Adam |
| Iteration | 10 | Scheduler | Cosine |
| Batch Size | 5 | DIM | 128 |
| Epoch | 10 | Alpha | 64 |
- Generating and Screening Humanoid Robot Design Options
An XY cross plot was used to generate Epoch model images under different weights and verify actual effects. Combined with the affinity score table, the XY cross plot was scored and screened. The study reports that the sixth Epoch model with weights from 0.6 to 0.8 produced the best effect. The XY cross plot scores showed that at Epoch 3 with Loss 0.089, the affinity scores were 9 at weight 0.2, 13 at 0.4, 13 at 0.6, 13 at 0.8 and 13 at 1.0. At Epoch 6 with Loss 0.075, the scores were 9 at weight 0.2, 9 at 0.4, 14 at 0.6, 14 at 0.8 and 13 at 1.0. At Epoch 9 with Loss 0.068, the scores were 9 at weight 0.2, 11 at 0.4, 12 at 0.6, 12 at 0.8 and 12 at 1.0.
| Epoch | Loss value | Weight 0.2 score | Weight 0.4 score | Weight 0.6 score | Weight 0.8 score | Weight 1.0 score |
|---|---|---|---|---|---|---|
| 3 | 0.089 | 9 | 13 | 13 | 13 | 13 |
| 6 | 0.075 | 9 | 9 | 14 | 14 | 13 |
| 9 | 0.068 | 9 | 11 | 12 | 12 | 12 |
The study emphasizes that affinity calls for personalization and diversity. By combining the trained model with different style models, it generated a large number of schemes with different styles and organized them into a style matrix. The style matrix was then evaluated with the affinity score table, allowing high-affinity schemes to be found quickly among many options. The reported style matrix showed that the first scheme in each style had the highest affinity score. For style 1, the five scheme scores were 15, 12, 10, 9 and 8. For style 2, the scores were 15, 14, 14, 14 and 14. For style 3, the scores were 15, 14, 14, 12 and 12. For style 4, the scores were 14, 12, 12, 9 and 9.
| Style | Scheme 1 | Scheme 2 | Scheme 3 | Scheme 4 | Scheme 5 |
|---|---|---|---|---|---|
| 1 | 15 | 12 | 10 | 9 | 8 |
| 2 | 15 | 14 | 14 | 14 | 14 |
| 3 | 15 | 14 | 14 | 12 | 12 |
| 4 | 14 | 12 | 12 | 9 | 9 |
- What the Results Mean for Humanoid Robot Design
The study concludes that Kansei engineering can provide a quantified basis for humanoid robot affinity design. By examining affinity, gentleness and liveliness, the research reveals appearance design features for an affinity-oriented humanoid robot and forms a precise design requirement and evaluation mechanism. This mechanism offers methodological guidance for humanoid robot design. With artificial intelligence-assisted design tools, the study trains a series of humanoid robot appearance design models based on affinity requirements. These models allow Stable Diffusion to generate a large number of design schemes with different styles in a short time, form a style matrix and quickly screen schemes.
The reported significance is twofold. First, the method can improve the efficiency and quality of humanoid robot appearance design. Second, it provides a new technical path for emotional design of robots and a reference for future affinity-oriented humanoid robot appearance design. The study suggests that such an approach can support the application and popularization of humanoid robots in daily life and commercial service scenarios and can help improve user acceptance and use intention. For humanoid robot developers, the work indicates that affinity should be treated as a design specification that can be measured, scored and iterated, rather than as a subjective preference added after engineering decisions are complete.
The study also acknowledges limitations. The questionnaire sample group was relatively concentrated, and the universality of the results requires validation with a broader range of users. Future work could expand toward cross-cultural research, diversified style generation and integrated research on appearance and interactive behavior. These directions are relevant for humanoid robot design because appearance does not operate in isolation: users’ affinity judgments may also be shaped by movement, speech, response timing and other behavioral cues. A more comprehensive design methodology could combine static appearance features with interaction behavior to support higher levels of humanoid robot affinity in multiple scenarios.
- Conclusion
The research presents a complete pipeline for affinity-oriented humanoid robot appearance exploration and Stable Diffusion model training. It begins with design-element abstraction and semantic differential questionnaires, uses statistical analysis to identify how head shape, eye shape, body proportion, material, surface treatment and color relate to affinity, gentleness and liveliness, and produces an affinity score table. That table then guides the selection and annotation of training samples, the iterative training of Dreambooth and Lora models, the evaluation of Loss values and XY cross plots, and the generation and screening of a style matrix. The reported result is a practical method for producing and evaluating humanoid robot appearance schemes with higher affinity in less time.
For the field of humanoid robot design, the study’s contribution lies in connecting affective design research with artificial intelligence generation tools through a quantified evaluation mechanism. The humanoid robot is treated not only as a technical system but also as an appearance-based social object whose form, material and color can influence acceptance and trust. The study’s findings suggest that a humanoid robot with a horizontal racetrack or circular head shape, vertical racetrack eyes, a slim juvenile body proportion, soft and textured fabric material, and bright, low-saturation warm or neutral color treatment may better support affinity. At the same time, the research shows that affinity-oriented humanoid robot design can be organized as an iterative process of measurement, sample creation, model training and scheme screening. This process may help design teams move from a small number of manually developed humanoid robot concepts to a larger, more diverse and systematically evaluated design space.