RoboLinker: A Diffusion-model-based Matching Clothing Generator Between Humans and Companion Robots

Jing Tang (唐晶), Jing Tang, Qing Xiao, Kunxu Du, Zaiqiao YearXiv 2025 · DOI · arXiv

Introduction

Emotional attachment is a cornerstone of Human-Robot Interaction (HRI). While significant research has focused on robot behavior and appearance in isolation, the concept of "robot fashion" has recently emerged as a powerful tool for personalization. Users often dress companion robots, such as the Sony AIBO, to foster a sense of companionship. However, this interaction is typically one-sided: the human styles the robot, but their own appearance remains disconnected from the machine.

RoboLinker addresses this gap by treating fashion as a "two-way medium." Drawing inspiration from matching outfits used by couples and families to signal social closeness, the system uses generative diffusion models to create human clothing that stylistically "rhymes" with a robot's attire. This approach transforms robot dressing from a solitary hobby into an act of "relational authorship," strengthening the emotional bond between humans and their robotic buddies.

Methodology

RoboLinker utilizes a diffusion-based architecture designed to bridge the aesthetic gap between robot form factors and human fashion. The system's workflow is built upon three technical pillars:

  1. Unified Tokenization: The system processes two distinct inputs—a textual style prompt from the user and a visual image of the robot's current clothing. These are mapped into a unified mathematical latent space, allowing the model to treat visual textures and linguistic descriptions as compatible data points.
  2. Diffusion Engine: At its core, RoboLinker employs a U-Net architecture that iteratively refines a human outfit design. This process ensures the resulting garment satisfies both the user's stylistic constraints and the visual cues (such as fabric, color, and pattern) extracted from the robot's attire.
  3. Interactive Co-Design: The system is implemented via an interactive interface where users can upload robot images and provide feedback. Automated recognition extracts features like silhouette and fabric type, while the user can refine the output in real-time through iterative prompting (e.g., "change the hemline").
The RoboLinker interactive interface facilitates a co-design cycle between the user and the generative model.

Datasets and Evaluation

The researchers evaluated RoboLinker using two primary robot categories to test its versatility across different morphologies:

  • Humanoid Robots (Wukong): A custom dataset of denim-themed clothing was created for the Wukong humanoid robot. This allowed the team to test how well the model could inherit specific stitching and fabric weights from a robot's garment into a human design.
  • Pet-like Robots (AIBO): The team collected a diverse dataset of AIBO robot dog images from social media (X/Twitter). This provided a range of "organic" fashion choices made by real-world users, testing the model's ability to translate accessories and pet-style clothing into stable human fashion.

Key Results and Insights

Qualitative analysis demonstrates that RoboLinker successfully generates coherent, wearable human outfits that mirror the "essence" of the robot's clothing. For humanoid robots, the system achieved high stylistic coherence, while for pet-like robots, it effectively translated small-scale accessories into full-scale human garments.

The RoboLinker system successfully translates a robot's garment (left) into a matching human fashion output (right).

Beyond technical matching, the authors identify "Fashion as Relational Authorship" as a key insight. The generative process stimulates "empathy points," where the resulting design reflects a deeper emotional resonance than simple color-matching. The interactive interface proved accessible for non-designers, allowing them to steer the generative model toward a desired aesthetic effectively.

Limitations and Future Work

While promising, the current iteration of RoboLinker has limitations. It is primarily optimized for specific robot forms and postures; future work will focus on adapting the model to a wider variety of robot shapes. Additionally, the authors plan to implement "reverse generation," allowing the system to design robot apparel based on a human's existing wardrobe. Finally, the team aims to bridge the gap between digital design and physical realization through integration with digital fabrication and smart tailoring technologies.

Summary updated

BibTeX
@misc{2025robolinker,
  author = {Jing Tang (唐晶) and Jing Tang and Qing Xiao and Kunxu Du and Zaiqiao Ye},
  title = {{RoboLinker: A Diffusion-model-based Matching Clothing Generator Between Humans and Companion Robots}},
  year = {2025},
  eprint = {2508.01165},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2508.01165},
}
Figure