Cybernetic Avatars: Embodiment & Autonomy

Updated 6 July 2026

Cybernetic Avatars (CAs) are embodied mediation systems that integrate human teleoperation, AI proxies, and hybrid control across digital and physical platforms.
They support dynamic switching between physical, AR, and VR bodies, enabling users to maintain a continuous identity while operating in multiple environments.
CA research explores autonomy, privacy, safety, and social impact, advancing methodologies for remote presence and secure, intuitive interaction.

Cybernetic Avatars (CAs) denote a heterogeneous class of embodied mediation systems in which a person, an AI proxy, or a hybrid of teleoperation and autonomy acts through robotic, virtual, or mixed bodies across physical and digital environments. In the literature surveyed here, the term covers distributed avatar systems that let a pilot switch among a physical robot avatar, an AR avatar, and a VR avatar; hybrid interaction robots or digital representations that combine autonomous capabilities with teleoperated control; virtual agents and robotic assistants envisioned for an avatar-symbiotic society; and social VR avatars that function as embodied digital agents whose spatially proximate actions can feel materially intimate (Yamazaki et al., 2022, Aymerich-Franch et al., 17 Jun 2025, Hafi et al., 17 Jul 2025, Guo et al., 2024). Rather than naming a single architecture, CA names a research space that joins embodiment, control, perception, communication, safety, and governance.

1. Conceptual scope and typologies

A recurring feature of CA research is definitional plurality. In the Meta Avatar Robot Cafe, a CA is effectively a distributed avatar system through which a remote “pilot” can switch among a physical robot avatar, an AR avatar, and a VR avatar while preserving one continuous social identity and work experience (Yamazaki et al., 2022). In service-robot research, CAs are defined as hybrid interaction systems that combine autonomous capabilities with teleoperated control, and they can appear either as physical robot avatars or as digital avatars (Aymerich-Franch et al., 17 Jun 2025). In the avatar-symbiotic-society literature, CAs are virtual agents and robotic assistants that enable individuals to overcome physical limitations and participate in daily life and work (Hafi et al., 17 Jul 2025).

This breadth is sharpened by control-oriented typologies. One strand distinguishes semi-autonomous CAs, which intermittently require human teleoperation and supervision, from fully autonomous CAs, which execute without human teleoperation during operation (Hafi et al., 17 Jul 2025). Another introduces the “independent avatar,” defined as an autonomous digital entity capable of managing social interactions on behalf of users in the Metaverse (Raptis et al., 16 Jan 2026). In virtual-performance research, the taxonomy becomes more fine-grained: avatar, guided actor, autonomous actor, and interactive-perceptive actor; or, in Plessiet’s two-axis formulation, Virtual Puppet, Virtual Actor, Virtual Mask, and Virtual Golem, distinguished by whether movement origin and decision are internal or external (Gagneré, 2024).

The literature also enlarges the scope of CA beyond telepresence and assistance. In social VR, avatars are treated as embodied digital agents through which users can approach, gesture, touch, insult, or otherwise affect others in ways that feel physically present, producing “embodied cyber threats” such as trash-talking, virtual “groping,” virtual harassment, and assault-like interactions (Guo et al., 2024). This suggests that CA is best understood not merely as an interface artifact, but as a controllable, socially consequential body schema instantiated in software, robotics, or both.

2. Embodiment architectures and body transfer

CA systems differ primarily in how they distribute embodiment across physical and virtual substrates. Some architectures make switching between bodies the central design principle. In the Meta Avatar Robot Cafe, the pilot operates from home through a tablet or PC application adapted to the pilot’s disability, works through a physical robot in a real cafe, and can switch to AR avatar operation or VR avatar operation in the cafe’s digital twin when the robot’s physical constraints become limiting (Yamazaki et al., 2022). In Avatar Robot Cafe DAWN ver.β, the same general logic is extended to parallel embodiment: one disabled worker may simultaneously manage multiple robotic avatars, creating multiple interaction loops in which the human relationship remains stable while the intermediary avatar changes (Barbareschi et al., 2023).

Form	Representative system	Salient property
Distributed multi-body CA	Meta Avatar Robot Cafe (Yamazaki et al., 2022)	One pilot switches among robot, AR, and VR bodies
Parallel robotic embodiment	Avatar Robot Cafe DAWN ver.β (Barbareschi et al., 2023)	One pilot can open up to four avatar interfaces simultaneously
Physical-metaverse CA	Avatarm (Villani et al., 2023)	Virtual hand is visible; robotic arm is hidden
Full-body android CA	Yui (Shinkawa et al., 6 Jun 2026)	Immersive teleoperation with full-body social signaling

Other CA architectures conceal the physical substrate in order to preserve avatar identity. Avatarm couples a virtual avatar hand to a hidden Sawyer robotic arm and uses a digital twin plus diminished reality so that the user sees the avatar as directly manipulating real objects while the robot remains visually absent (Villani et al., 2023). The system’s manipulation pipeline maps virtual hand action to virtual object dynamics, then to real object dynamics, and finally to robot joint commands. By contrast, the android CA “Yui” foregrounds embodiment rather than hiding it: the 2023 head-unit system integrates 28 actuation points, natural gaze control via HMD eye tracking, stereo microphones in the ears, and ROS2-based motor control so that the operator can transmit gaze, facial emotion, and speech-related mouth movements in real time (Nakajima et al., 2023). The 2026 full-body version extends this to a 55-degrees of freedom platform with arms, wrists, hands, waist, mobility, and two operator modes—immersive HMD-based control and webcam-based desktop control (Shinkawa et al., 6 Jun 2026).

Across these systems, embodiment is not fixed to one body. It can be sequential, simultaneous, partially hidden, or expanded from a head unit to a full-body android. A plausible implication is that CA research treats “body” less as a singular object than as a dynamically allocable resource for presence, action, and social continuity.

3. Autonomy, delegation, and control regimes

The control problem in CA research spans direct teleoperation, mixed-initiative systems, and full autonomy. In the Avatar Land demonstration, fully autonomous robotic CAs executed daily object retrieval and delivery in a replicated home environment, including object identification, navigation, grasping, carrying, and disposal handling (Hafi et al., 17 Jul 2025). The paper’s most explicit formal element is the exophora-resolution probability for target-object identification, computed as $p = p_1 p_2 p_3$ , where $p_1$ is related to the user’s pointing gesture, $p_2$ to demonstrative words, and $p_3$ to the object-category distribution in the environment (Hafi et al., 17 Jul 2025). This formulation situates CA autonomy not only in motion planning but also in reference resolution and instruction grounding.

A more abstract autonomy model appears in metaverse social-network optimization. The independent-avatar framework models avatar-mediated interaction through a social-presence conversion factor $\beta_v$ and a debriefing cost $z_v = \gamma \cdot y_v$ , where $\gamma = c \cdot \delta$ (Raptis et al., 16 Jan 2026). With the paper’s grounded values, $\beta \approx 1.29$ and $\gamma \approx 0.63$ , the system falls in the regime where avatar delegation is beneficial because $\gamma \le 1/\beta$ (Raptis et al., 16 Jan 2026). Once discrete time slots, conflict graphs, deadlines, and penalties are introduced, the full scheduling problem becomes NP-hard by reduction from CLIQUE, and the authors therefore introduce a heuristic based on sorted sequential request scheduling (Raptis et al., 16 Jan 2026). In simulations over 10,000 ego networks, avatar mediation reduced social cost, and the relative improvement remained substantial under higher conflict density, decreasing from about 95% to about 75% as conflict increased (Raptis et al., 16 Jan 2026).

Virtual-performance research addresses autonomy differently, as a staged effect. “Directing Avatars in Live Performances” argues that avatars can move from externally controlled vehicles toward entities with internalized mechanisms of movement, perception, and decision, but the operative result is a “simulacrum of autonomy,” not genuine philosophical self-rule (Gagneré, 2024). The technical means include root motion, a capsule-based locomotion model, finite state machines, pathfinding on NavMesh, Behavior Trees, and the Salient-Idle Player used in The Shadow to alternate expressive gestures with waiting/listening loops (Gagneré, 2024). This establishes a distinct CA regime: autonomy as cybernetic orchestration whose credibility depends on timing, internal action-selection logic, and audience interpretation.

A separate delegation regime appears in AI-mediated communication. ProxyMe is a VR prototype in which the user speaks first, Whisper transcribes the utterance, Llama-3.1-8B modifies it, and IndexTTS renders the transformed utterance either in a cloned version of the user’s voice or a neutral synthetic voice (Zhang et al., 6 Mar 2026). The system is explicitly designed to study partial delegation rather than full automation, with research questions centered on agency, authorship, and self-identification. Its measured end-to-end latency is about 11.6 seconds, which the authors identify as a serious obstacle for self-extension (Zhang et al., 6 Mar 2026).

4. Representation, expression, and the phenomenology of presence

A major CA line of work concerns photorealistic telepresence. Pixel Codec Avatars (PiCA) redesign the rendering side of codec avatars so that decoding scales with visible pixels rather than a full fixed texture, using a dense geometry decoder, a fully convolutional expression decoder, and a rendering-adaptive per-pixel SIREN-style decoder (Ma et al., 2021). The model has 5.47M parameters versus 19.08M for the baseline, reduces DSP time from 44.76 ms to 2.16 ms, reaches 90 FPS for single-avatar rendering, and renders five avatars simultaneously at about 50 FPS on one Oculus Quest 2 (Ma et al., 2021). Modular Codec Avatars (MCA) pursue a complementary goal: replacing holistic face codes with per-camera modular encoders and modulated adaptive blending, yielding better expressiveness and robustness for headset-driven telepresence. Reported aggregate improvements include MAE 6.54 to 6.17, RMSE 5.81 to 5.48, geometry RMSE 1.37 to 1.17, texture RMSE 2.72 to 2.44, and runtime of 21.6 ms per frame (Chu et al., 2020).

Expression fidelity does not straightforwardly translate into comfort. In a study of four autonomous ECAs and two human-controlled, face-tracked avatars, the two avatars were judged more human-like than the autonomous agents, but also more eerie (Thaler et al., 2021). Across 215 participants, the humanness–eeriness relationship was positive and substantial, with Pearson correlation $p_1$ 0, linear-regression $p_1$ 1, and significant differences such as Alice vs. Sophie on humanness ( $p_1$ 2) and eeriness ( $p_1$ 3) (Thaler et al., 2021). The paper therefore supports the Uncanny Valley interpretation that increased perceived humanness can increase eeriness rather than acceptance.

Other CA systems target expressivity through embodied mechanics and multimodal feedback rather than neural rendering alone. The head-unit version of Yui can reproduce seven basic emotions, uses stereo cameras and stereo ear microphones, and enables operators to infer sound direction from binaural cues while transmitting gaze and facial movement to the interlocutor (Nakajima et al., 2023). In teleoperation trials, the system supported eye–neck coordination and nonverbal greeting behaviors including jaw motion, mouth-corner movement, cheek motion, and blinking (Nakajima et al., 2023). The full-body Yui deployment extends this to a socially deployable android form that interlocutors rated positively in terms of human likeness and the transmission of emotions and intentions (Shinkawa et al., 6 Jun 2026).

AI-driven persona avatars move the discussion from facial fidelity to identity fidelity. “Digital Avatars: Framework Development and Their Evaluation” proposes an end-to-end pipeline from STT to LLM prompting, TTS, talking-face synthesis, and real-time streaming, with a “show don’t tell” prompting strategy that conditions persona-specific humor, tone, and style (Rupprecht et al., 2024). In the paper’s Crowd Vote evaluation, 58.3% of judges favored the authors’ avatar as the most humorous, and the authors report higher authenticity and favorability than baseline LLMs, Character.ai, and even the corresponding real-world persona responses in their evaluation setting, which used 17 questions from the first 2020 presidential debate for authenticity (Rupprecht et al., 2024). ProxyMe addresses the inverse problem: not how convincing the avatar is to observers, but whether AI-modified speech delivered through an embodied avatar and a cloned voice is still experienced as “me” speaking (Zhang et al., 6 Mar 2026).

5. Privacy, traceability, and safety

CA systems inherit the privacy problems of networked identity systems and add new ones tied to embodiment and autonomy. In mobile social metaverses, personalized 3D avatars generated from prompts, sensing data, and face-related attributes can create both profile identity leakage and persistent digital identity exposure (Su et al., 2024). To address this, the paper introduces avatar pseudonyms, the Privacy of Personalized Avatars (PoPA) metric, a Stackelberg-game formulation for pseudonym distribution, and PPO-based deep reinforcement learning under incomplete information (Su et al., 2024). In the reported simulations, the DRL method outperformed a random baseline by 27.2% and a greedy baseline by 55.7%, while also achieving higher PoPA than DreamFace (Su et al., 2024).

Traceability research addresses a different threat model: impersonation by AI-driven avatars. “Traceable AI-driven Avatars Using Multi-factors of Physical World and Metaverse” binds avatar identity to the manipulator’s iris feature and the AI proxy’s public key, and uses a chameleon proxy signature to support legitimate delegation while preserving accountability (Yang et al., 2024). The resulting protocols reportedly complete user login, avatar delegation, mutual authentication, and avatar tracing in about 1s, and avatar tracing itself in less than 500 ms, while the security analysis claims unforgeability and defense against false accusation (Yang et al., 2024). Here the CA is not only an embodied agent but a cryptographically accountable one.

Safety concerns are equally embodied. In social VR, the problem is not only identity theft but avatar-mediated abuse. “Moderating Embodied Cyber Threats Using Generative AI” argues that avatars create a channel for proximal, body-like interaction and therefore for intensified cyber threats such as trash-talking, virtual “groping,” virtual harassment, and assault-like interactions (Guo et al., 2024). The paper does not introduce a formal moderation algorithm, but proposes a conceptual mitigation agenda in which GenAI assists real-time moderation by drawing on broad training data to understand human behavior, using chain-of-thought reasoning for contextual analysis, and improving from human feedback; it simultaneously warns that the same GenAI can generate more convincing avatars or environments and thereby amplify deception and abuse (Guo et al., 2024). Related industrial HRC work reports strong discomfort with recording or sharing worker data and with monitoring emotion, attention, or performance, indicating that CA governance must also address transparency, surveillance boundaries, and trust in deployment settings (Klein et al., 10 Jun 2025).

Accessibility and disability support are among the most developed application domains. The Meta Avatar Robot Cafe is explicitly designed to let people with disabilities who have difficulty going out switch among physical and virtual bodies, work remotely, and communicate “presence and warmth” to others; one bedridden participant reported having “sat in a chair for the first time in several years” through avatar control (Yamazaki et al., 2022). In the parallel-control cafe study, disabled workers remotely embodied multiple OriHime-based avatars at once, opening up to four avatar interfaces, following customer journeys across greeting, reservation checking, table interaction, and drink delivery, and increasing their agency in managing service (Barbareschi et al., 2023). The paper’s title phrase—“I am both here and there”—captures the reported experience of simultaneous embodied locations (Barbareschi et al., 2023).

Industrial CA work shifts emphasis from disability access to human-centric production. In a one-day workshop with three employees from a German manufacturing company already using cobots, the avatar was mounted on a screen near the cobot base and was interpreted variously as communicator, motivator, explainer, colleague, or even potential impersonation of the worker (Klein et al., 10 Jun 2025). Participants appreciated personal communication, especially when the avatar addressed the company directly, but rejected messages perceived as rude or pressuring. Desired future behaviors included break reminders, job rotation suggestions, early error detection and correction guidance, safety reminders, and adaptation to workload, mood, and attention (Klein et al., 10 Jun 2025).

Service-sector acceptance has been studied at scale. In Dubai, a survey of 1,001 participants found higher acceptance for physical robot avatars (67.3% agree) than for digital avatars (56.9% agree), with a significant Wilcoxon Signed-Rank result of $p_1$ 4 (Aymerich-Franch et al., 17 Jun 2025). Among appearances, highly anthropomorphic robotic-looking avatars ranked highest at 61.1% agree, followed by cartoonish designs at 53.3% and ultra-realistic androids at 50.4%; animal-like forms ranked lowest at 39.0% (Aymerich-Franch et al., 17 Jun 2025). The most accepted deployment settings were shopping malls (74.5%), airports (69.6%), train and metro stations (68.9%), and museums (69.1%), whereas hospitals and clinics (37.4%), rehabilitation centers (35.9%), and nursing homes (32.3%) received markedly lower support (Aymerich-Franch et al., 17 Jun 2025). The most accepted tasks were providing information to customers (72.5%) and providing guidance or directions (71.7%) (Aymerich-Franch et al., 17 Jun 2025).

Public evaluation of fully autonomous CAs points to a different constraint: reliability. At Avatar Land, 2,285 survey respondents were recorded overall, with 333 participants specifically engaging the fully autonomous daily-life support CA demonstration (Hafi et al., 17 Jul 2025). Among that subset, 39.3% answered “Very likely to use” and 35.4% answered “Use if conditions are right,” yielding 74.7% positive attitude toward use; the most common expected contexts were daily life (47.8%) and work (32.5%) (Hafi et al., 17 Jul 2025). The dominant reason for reluctance was “Could not use well” at 37.5% of negative responses, whereas cost and human-like interaction were not dominant concerns (Hafi et al., 17 Jul 2025). This aligns with field deployment results from the full-body Yui system: during Expo 2025, two units accumulated approximately 1131 h of operation over 113 exhibition days, demonstrating operational feasibility but also maintenance burden, battery degradation, and the practical need for hybrid autonomous playback plus teleoperation (Shinkawa et al., 6 Jun 2026). In public interaction studies, both operators and interlocutors reported positive impressions of co-presence and willingness to use the system, while interlocutors also rated it positively on human likeness and the transmission of emotions and intentions (Shinkawa et al., 6 Jun 2026).

Taken together, these evaluations indicate that CA adoption is shaped less by a single variable than by a coupled set of conditions: embodiment quality, controllability, reliability, privacy guarantees, contextual fit, and the social meaning of the role delegated to the avatar. The literature therefore presents CAs not as a unified device category, but as an evolving infrastructure for remote action, mediated sociality, and distributed personhood.