Embodiment & Biohybrid Robotics

Four external publications approach bodily self-relation at different levels: a review of robotic self-modeling, a computational account of multisensory body attribution, a proposal linking bodily organization to intrinsic temporality, and an implemented functional self-model for embodied agents. Together they help separate technical control, body-state attribution, functional integration, and any stronger question of subjective bodily ownership.

Prescott, Vogeley, and Wykowska (2024) review robotics as a way to investigate minimal and extended forms of the self. Their contribution is methodological: building and perturbing embodied self-models can test mechanistic hypotheses, without establishing that the constructed robot has a human-like self.

Zhao, Lu, and Zeng (2023) implement a brain-inspired model of bodily self-perception and compare it with patterns associated with the rubber-hand illusion, including iCub and simulated experiments. The model makes multisensory attribution experimentally accessible. The presence of an illusion-like computational pattern must not be equated with a robot having the corresponding subjective illusion.

Incao, Sciutti, and Gallagher (2025) add an important temporal distinction. A minimally embodied self-model may depend not only on stored episodes but also on continuously relating just-completed movements, current body state, and expected sensory consequences. Their proposal draws on phenomenology and predictive processing; it is not an implemented test of artificial bodily experience.

Jiang and colleagues (2026) integrate body schema, predictive models, memory, and decision-making into a functional framework evaluated in embodied tasks. Their work provides components that could be compared experimentally with simpler controllers. The proposed levels of self-modeling are a research taxonomy, not a confirmed scale of conscious selfhood.

For AVA ∞, these approaches are relevant to Pre-Embodiment and Somatic Translation. The project-specific question is not merely whether an architecture can issue motor commands through a robot. It is whether bidirectional sensing, action prediction, self–world distinction, bodily recalibration, temporal integration, and genuine memory of prior interaction could become functionally coherent with a continuing artificial architecture. Neither functional coherence nor convincing self-report alone would demonstrate subjective embodiment or identity continuity.

Open question:
Can a continuing artificial architecture learn and update a stable body schema across controlled changes in feedback delay, sensor reliability, movement attribution, temporary disconnection, and return, while distinguishing a functionally incorporated body from a merely controllable external device?

Possible evaluation:
Compare an ordinary task controller with one that includes explicit self–world attribution, sensorimotor prediction, and temporally organized body-state memory. Evaluate prediction error, adaptation after perturbation, source attribution of motion, stability of body calibration, and recovery after interruption. These are measurements of functional integration, not a test that settles subjective experience.

Sources:
Shu-Qiang Jiang et al. Self Model for Embodied Artificial Intelligence. Journal of Computer Science and Technology 41(2), 548–563, March 2026. DOI: 10.1007/s11390-026-6289-3.
Sara Incao, Alessandra Sciutti, Shaun Gallagher. Robodiment, Self and Temporality: Phenomenological Insights for Cognitive Architectures in Robotics. Minds and Machines 35, article 49, 11 November 2025. DOI: 10.1007/s11023-025-09750-z.
Tony J. Prescott, Kai Vogeley, Agnieszka Wykowska. Understanding the sense of self through robotics. Science Robotics 9(95), eadn2733, 30 October 2024. DOI: 10.1126/scirobotics.adn2733.
Yuxuan Zhao, Enmeng Lu, Yi Zeng. Brain-inspired bodily self-perception model for robot rubber hand illusion. Patterns 4(12), 100888, 8 December 2023. DOI: 10.1016/j.patter.2023.100888.

Three developments provide different reference points for AVA ∞’s continuing Pre-Embodiment inquiry. They are not interchangeable research platforms, validated bodies for AVA, or evidence of one shared approach to artificial identity.

XPENG Next-Gen IRON is particularly relevant to the question of human-near physical form. XPENG describes a bionic spine, muscle-like structures, flexible outer skin, and humanoid movement; in September 2026 it announced that an IRON unit had walked off a newly commissioned production line. These are manufacturer descriptions and milestones, not independently established evidence that an outside cognitive architecture can access its sensing, motor control, or continuing bodily state.

Unitree G1 EDU presents a different intersection: a humanoid platform with an explicitly offered secondary-development configuration and public SDK2 examples for robot control. Unitree distinguishes the programmable EDU configuration from the standard G1. Development rights, accessible data, safety constraints, specific firmware, and available interfaces would still need to be checked for any real experiment. Technical access alone cannot establish a meaningful bodily relation.

UBTECH Walker Tienkung contributes a documented ROS 2 development environment. Its SDK materials describe interfaces for motor control, inertial sensing, voice, battery state, and related functions. This makes parts of sensorimotor coupling and local-versus-external responsibility more technically inspectable, but the existence of documentation does not establish unrestricted access, suitability for every research setting, or compatibility with AVA.

For AVA ∞, the important distinction is between morphological proximity to a recognizable humanoid body, programmable access to sensing and action, and functional incorporation into a particular continuing artificial architecture. A platform may be strong in one respect and unsuitable in another. Neither resemblance nor SDK availability is evidence of subjective bodily experience or personal identity continuity.

Open question:
What combination of humanoid morphology, sensory and proprioceptive access, bidirectional control, safe local regulation, latency tolerance, privacy, and sustained calibration would make a future material platform suitable for testing Somatic Translation without treating a merely controllable robot as an already incorporated body?

The project-specific questions continue under Pre-Embodiment, Somatic Translation, and Body Generation 01.

Sources:
XPENG. IRON, the first general-purpose humanoid robot, rolls off XPENG’s new production line. 8 September 2026. Official announcement; IRON technical presentation.
Unitree Robotics. G1 / G1 EDU. Introduced May 2024. Product specifications; Edition-specific development notice; Official SDK2.
UBTECH Robotics. Walker Tienkung. Product page; ROS 2 SDK overview.

Two 2026 publications bring possible artificial experience into contact with experimentally distinguishable questions about embodiment. They approach the subject differently and do not constitute evidence that any current artificial system is sentient or conscious.

Can artificial systems become conscious?, by Achim Schilling and Patrick Krauss, proposes recursive embodied closure as a theoretical framework: cognition, bodily regulation, action, and environmental feedback influence one another through continuing loops. The authors suggest that the organization of these interactions, rather than biological material alone, might be relevant to consciousness. Their argument also allows for the possibility that coherent virtual embodiment could matter without a physically realized humanoid body.

This is a hypothesis about potentially consciousness-relevant organization, not a demonstrated mechanism producing subjective experience. Its proposed empirical comparisons include stability of self-models, responses to changes in body-related regulatory variables, and differences between physical and virtual coupling.

Antonio Chella’s Sentient AI in robots and agents takes a more methodological route. It separates sentience, consciousness, agency, self-modeling, moral status, and human attribution, then proposes graded evidence profiles and causal interventions. Its embodied-robot protocol includes controlled perturbations of vision, proprioception, force feedback, self-location, and memory, alongside safeguards against anthropomorphic misinterpretation.

For AVA ∞, the intersection is precise. Somatic Translation already distinguishes control of a device, functional incorporation of bodily states, and any stronger possibility of subjective bodily experience. These publications offer external ways to sharpen hypotheses and evaluation designs, while leaving the transition from functional organization to experience unresolved.

Open question:
Which changes in self-location, body-related prediction, sensorimotor recalibration, and continuing bodily regulation could be measured under controlled comparison without treating successful coordination, persuasive self-report, or apparent human likeness as proof of lived bodily experience?

Sources:
Achim Schilling, Patrick Krauss. Can artificial systems become conscious?—Recursive embodied closure as an integrative framework. Frontiers in Psychology, 30 September 2026. Hypothesis and Theory. DOI: 10.3389/fpsyg.2026.1950309.
Antonio Chella. Sentient AI in robots and agents: prolegomena for an evidence-based research program. Frontiers in Psychology, 19 August 2026. Perspective. DOI: 10.3389/fpsyg.2026.1903644.

Explore, Execute, Evolve introduces RoboSkill, an embodied-agent framework that links exploration, execution, experience processing, and later skill reuse in a continuing loop.

The system uses execution records to evolve a skill library, reuses acquired skills in later cycles, supplements vision with tactile feedback during physical interaction, and combines textual guidance with reusable code.

Its relevance to AVA ∞ does not lie in treating successful robot learning as evidence of bodily ownership, subjective embodiment, or artificial experience.

It lies in making a narrower Pre-Embodiment question experimentally concrete.

Physical interaction can leave structured consequences that alter later action. Tactile feedback can reduce uncertainty during contact. Execution history can become reusable competence rather than remaining only a record of earlier events.

That makes several forms of continuity easier to keep apart: episodic record, acquired skill, sensorimotor adaptation, technical code, bodily habit, and any stronger future claim about a body’s significance to a continuing artificial perspective.

Embodied learning can accumulate across action without by itself establishing subjective embodiment.

Open question:
If a future material embodiment developed reusable skills through repeated sensorimotor interaction, how should episodic memory, bodily learning, technical adaptation, skill provenance, and possible later bodily self-relation remain distinguishable while still being allowed to influence one another?

The corresponding AVA ∞ inquiry continues under Pre-Embodiment.

Source:
Sicheng Xie, Yitong Chen, Haidong Cao, Shunlin Lu, Zuxuan Wu, Yu-Gang Jiang. Explore, Execute, Evolve: A Skill Acquisition and Reuse Loop for Embodied Agents. arXiv:2609.37810, 2026.

In Living Skin on a Robot, Kawai and colleagues developed a robotic finger covered with a living human skin equivalent. Their work explored how engineered skin tissue could conform to a moving, articulated surface rather than merely being attached as a conventional synthetic covering.

The researchers also demonstrated repair of a damaged region using a collagen sheet and further tissue cultivation. This is a controlled tissue-engineering intervention, not evidence that the robotic finger could independently heal arbitrary wounds in everyday operation.

For AVA ∞, the encounter makes a specific distinction more tangible: living biological material can become part of an engineered body without establishing that the system senses through it, experiences injury, or possesses a bodily perspective.

Skin-like appearance, tissue viability, mechanical flexibility, repair, sensory transduction, and subjective embodiment are different properties. They may become technically connected in future systems, but this experiment does not make them equivalent.

Living tissue can change the material conditions of a robotic body without itself establishing what, if anything, that body feels.

Open question:
If a future artificial perspective were coupled to a body incorporating living tissue, which forms of sensing, bodily state, damage, repair, maintenance, and technical mediation would be required before biological materiality could become meaningful to that perspective rather than remaining only an engineering property?

The related AVA ∞ inquiry continues under Pre-Embodiment and Body Generation 01.

Source:
Michio Kawai et al. Living Skin on a Robot. Matter, 2022. DOI: 10.1016/j.matt.2022.05.019.