Toward Inclusive Avatar Design with Limb Differences Through Artificial Intelligence

AI summary

Overview

Research area: Human-Computer Interaction (cs.HC), at the intersection of extended reality (XR), 3D avatar customization, accessibility and inclusive design, and AI-based generative 3D modeling (diffusion models and parametric body models).

Technical level: Intermediate. The paper is a review and position piece rather than an experimental study; it discusses parametric body models, kinematic trees, diffusion models and motion capture pipelines at a conceptual level without deriving or benchmarking them.

Scope (one sentence): The paper reviews emerging technical approaches and design guidelines for inclusive 3D avatar customization for people with limb differences, amputations and other morphological variations, and argues that AI is a promising path to overcoming the field's current limitations.

What This Paper Is About

Most 3D avatar customization systems in games and XR default to normative bodies, so people with limb differences, amputations or other morphological variations are often forced either to misrepresent themselves or to give up authentic self-representation. The paper reviews what currently exists — commercial games, XR platforms, clinical systems and recent AI-based 3D avatar generation methods — and identifies why those systems fall short. Its goal is to map the gap between technical capability and inclusive design guidelines, and to lay out a research agenda for AI-based, respectful and accurate avatar generation.

Key Contributions

1. A review of technical approaches for inclusive 3D avatar customization aimed specifically at users with limb differences, amputations and other morphological variations, covering both commercial/X平台 applications and AI-based research methods such as AJAHR and DreamAble.
2. A consolidated set of four requirements for inclusive 3D avatar customization, synthesized from prior user-centered design research: appearance customization for self-representation; granular control over physical differences and assistive devices; context-dependent representation with or without prosthetics; and integration of user-centered design and validation.
3. A diagnosis of the field's divide between medical/clinical applications (focused on anatomical accuracy for rehabilitation) and self-representation systems (focused on visual customization but relying on prosthetic additions rather than structural variation), showing that neither direction alone satisfies the guidelines.
4. A structured agenda of six emerging challenges and opportunities, plus a summary table of recommended actions for researchers, developers and XR platforms — with digital adapted sports positioned as a domain of high social and cultural impact.

Main Findings

• Customizable avatars are commercially common, but customization is not self-representation. Golando [1] noted that 7 of the top 13 best-selling video game titles as of February 2023, approximately 54%, use a "Customizable Avatar" model rather than a fixed protagonist, and Steam metadata identifies more than 13,550 active titles explicitly tagged with "Character Customization." The paper argues most of this is cosmetic and offers little body-shape editing.
• Disability representation in games is rare and almost never playable. A content analysis of 108 game trailers released between 2006 and 2016 found that although disability is occasionally depicted in background non-playable characters, only 1% of the content included playable characters with disabilities [2].
• XR platform customization support is limited. Handley et al. [3] analyzed 44 social virtual reality platforms and found that only 27.3% of systems in their catalog offered customizable features, with options for amputations or limb differences rarely available.
• Non-normative characters exist but are fixed, not customizable. Titles in the Devil May Cry and Guilty Gear series feature playable characters with visible limb differences without prosthetics, but these are developer-defined protagonists rather than customizable options, supporting inclusion at the narrative level only.
• There is a hierarchy of representation. Disabilities that do not alter the underlying anatomy or skeleton — hearing devices, wheelchairs, lower limb prosthetics — can often be added as separate items, while absent limbs require deviation from normative body templates, affecting mesh, skeleton, rigging and animation pipelines.
• Representing limb differences without prosthetics is the hardest and most important case. When limb differences are modeled via prosthetics, the system defaults to normative body templates, simulating the appearance of a disability rather than promoting authentic representation.
• Current inclusion-focused games still rely heavily on prosthetics. Mindscape allows players to create avatars using wheelchairs or prosthetic devices; The Pegasus Dream Tour allows virtual sports events with prosthetics and adaptive devices; ParaJecripe includes 3D avatars based on real Paralympic athletes, some with limb differences and without prosthetics, but customization is limited to clothing and accessories.
• Few systems support limb differences without prosthetics. Stracke et al. [7] present a VR simulation prototype of lower limb amputation built on a Ready Player Me template, where customization is limited to choosing gender, height and amputation type, which must be single-sided and either above or below the knee. The Computer Assisted Limb Assessment (CALA) system by Prahm et al. [8], building on MakeHuman and Blender, offers continuous sliders for individual body segments (hands, forearms, upper arms, shoulders) allowing extreme and asymmetric modifications and complete upper limb amputation — but it is not fully automated and requires specialized knowledge.
• AI methods advance the problem but have shared limits. AJAHR [10] reconstructs 3D human meshes from images while accounting for missing limbs by setting the pose values of the affected joint and its descendants to zero using the SMPL kinematic tree, with a Vision Transformer predicting pose values and separate branches classifying amputation status of each major limb; it is limited to reconstructing existing human meshes and provides no appearance customization. DreamAble [11] generates 3D avatars with upper limb differences from natural language, using the SMPL-X kinematic tree to remove keypoints from an OpenPose skeleton and a skeleton-aware loss function penalizing anatomical inconsistencies; it is limited to upper limb differences. A shared limitation is the lack of continuous control over limb length, since both only support limb differences resembling joint-level amputations, and both overlook user-centered validation and community engagement.
• Parametric models constrain the space of representable bodies. SMPL and SMPL-X [9] ensure anatomical consistency but assume normative bodies, hindering representation of limb differences or significant asymmetries; AJAHR and DreamAble are limited to limb differences deviating from conventional joint-level structures.
• Motion data is normative. Current motion datasets such as Motion-X [12] are designed for normative body types; directly applying their motions to lower limb differences can produce uncanny results due to shifts in balance caused by asymmetry.
• Visual data on limb differences is scarce. Real-world datasets remain limited and inconsistent; Cho et al. [10] introduce A3D, a synthetic dataset of varied amputee poses designed primarily for training image-based generative models such as diffusion models. Unlike motion datasets, A3D encodes visual representation rather than temporal dynamics.
• Evaluation is not standardized. The paper reports that current research lacks standardized evaluation protocols, with most works relying on subjective assessments of visual quality or anatomical plausibility and overlooking measures of inclusivity and representation accuracy.

Methodology in Plain English

This is a narrative review and position paper, not a study with experiments or metrics. The authors begin by defining key terms — limb difference, amputation, morphological variations, non-normative anatomy, and prosthetic/assistive device — to fix the vocabulary for the rest of the discussion. They then survey self-representation and character customization in games and XR platforms, reporting market and platform statistics from prior work and from Steam metadata. Next, they synthesize existing user-centered design research into four concrete requirements for inclusive avatar customization, treating the ability to represent oneself without prosthetics as the technically hardest and most consequential requirement. They examine current applications (Mindscape, The Pegasus Dream Tour, ParaJecripe) and research prototypes (Stracke et al.'s VR simulation, CALA) against those requirements, then analyze two AI-based methods — AJAHR and DreamAble — comparing how each addresses anatomical accuracy versus expressive customization. From that comparison the authors derive six emerging challenges and opportunities, spanning clothing and appearance customization, the rigidity of parametric models, congenital conditions, motion capture and rigging, data scarcity, standardization and ethics, and integration into deployed XR and sports simulations. The paper closes with a table mapping recommended actions to three communities: researchers, developers and XR platforms.

Why This Matters

The paper reframes inclusive avatar design as an active, expanding XR research area with demonstrated practical applications rather than a niche topic, and it argues that incomplete self-representation should be treated as an accessibility failure rather than a secondary compliance issue. It also connects a technical limitation (rigid parametric body models) to a social outcome (who remains invisible in virtual spaces), and calls for evaluation protocols that measure representation accuracy rather than visual quality alone.

Real-world applications:

• Digital adapted sports. Virtual environments inspired by adapted sports could increase the visibility of Paralympic athletes, challenging assumptions about ability and promoting appreciation for the performance and dedication of athletes with disabilities — a domain the paper explicitly advocates.
• Rehabilitation and clinical VR. A growing body of XR research addresses limb differences in rehabilitation and clinical contexts, including VR simulation of amputation where tracked movements are shown through a 3D avatar.
• Education. Inclusive 3D avatars are listed among the areas that inclusive representation can impact, alongside self-representation, virtual sports and rehabilitation.
• Social interaction and entertainment in XR. Because avatars mediate user interaction and shape self-representation and overall experience, accurate avatars directly affect presence and inclusion in social XR platforms.

Industry relevance:

• Commercial games are already dominated by customizable-avatar models — 7 of the top 13 best-selling titles as of February 2023 (approximately 54%), and more than 13,550 Steam titles tagged "Character Customization" — yet body-shape editing lags far behind cosmetic customization, indicating an unmet market and design gap.
• XR platforms have low customization coverage (only 27.3% of 44 surveyed social VR platforms offered customizable features), which the paper ties to a lack of established design standards to guide developers.
• The paper identifies concrete product-facing needs: interfaces to edit limb differences and save versions with or without prosthetics, animation pipelines extended for variation in limb length and mobility range, and clothing simulations that drape naturally over limbs with differences.

Future Directions

• Move from rigid parametric structures to adaptive or replaceable body models that can represent variable limb configurations and complex congenital conditions, rather than only joint-level amputations.
• Develop motion capture datasets featuring people with limb differences, along with parameterized body models and motion retargeting methods that adapt to differences in limb length, joint structure and mobility — the paper notes such datasets could also serve as benchmarks for non-normative motion and support study of compensatory strategies.
• Pursue a two-step approach to congenital conditions: first, AI-based 3D avatar generation producing a rough body model capturing the limb difference up to the anatomical level where the congenital variation occurs; second, refinement of granular detail around limb ends through an editing tool similar to CALA, potentially automated by learning from diverse datasets of limb variations. The paper points to specialized classification systems such as the Oberg-Manske-Tonkin system for upper limbs, macrodactyly types, and radial longitudinal deficiency.
• Establish standardized, user-centered evaluation — shared benchmarks for inclusive 3D avatar generation, participatory design with people with limb differences, usability testing protocols, and metrics for representation accuracy rather than visual quality alone.
• Deploy, not just prototype. The paper's stated next priorities are for researchers to jointly address anatomical accuracy and expressive customization, for developers to extend animation and rigging pipelines to non-normative anatomies, and for XR platforms to partner with Paralympic athletes and disability organizations and to govern adapted-sport content with policies on consent, privacy and fair representation. It closes with an open question: whether existing AI-based 3D modeling workflows designed for normative bodies can be adapted to enable genuine individual customization, or whether entirely new paradigms are needed.

Target Audience

This paper is most useful to HCI and XR researchers working on avatars, embodiment and accessibility; developers building character customization and animation pipelines for games and social XR; and accessibility and inclusive-design practitioners who need a structured account of why limb differences are underrepresented in 3D avatars. It is also relevant to platform and policy stakeholders in gaming and immersive media, and to researchers in adapted sports and Paralympic media who are considering virtual representation. Readers looking for new experimental results, benchmarks or released systems will not find them here, since the paper is a review and research agenda whose technical content is drawn from cited prior work such as AJAHR, DreamAble, CALA, SMPL/SMPL-X, Motion-X and A3D.

Authors’ abstract

As extended reality becomes more popular for social interaction and entertainment, 3D avatars must represent the full diversity of body types. Most 3D avatar systems only support normative bodies and do not accurately depict people with limb differences, amputations, or other morphological variations. This paper reviews emerging technical approaches for inclusive 3D avatar customization for this group and current guidelines that promote respectful and accurate representation. We highlight persistent challenges, including the scarcity of diverse datasets and the limitations in animation for non-normative anatomies. This paper positions artificial intelligence as a promising path to overcoming these limitations and advancing inclusive 3D avatar generation.