Trans-Embodiment: Adaptive Robotics and Identity

• Trans-embodiment is the ability of a single controller to operate without fine-tuning across varied robot forms and digital avatars, unifying robotics and HCI.
• It leverages morphology-aware neural policies, curriculum randomization, and shared latent spaces to achieve robust zero-shot skill transfer.
• Empirical evidence shows high success in locomotion and manipulation tasks, while highlighting challenges in handling out-of-distribution embodiments.

Trans-embodiment refers to the capacity of a policy, agent, or interaction system to operate seamlessly across widely differing physical forms (“embodiments”). In robotics, trans-embodiment typically signifies a single controller or representation being deployed zero-shot, without fine-tuning, across agents with radically distinct morphologies, action spaces, or sensorimotor characteristics. In cognitive and HCI contexts, it can reference the migration or mapping of personality, identity, or agency across time, space, or substrates, often imbuing continuity of presence or function. This unifying principle underlies numerous advances in generalist robotics, cross-morphology reinforcement learning, skill imitation, and agent design.

1. Formal Definitions and Theoretical Foundations

Trans-embodiment, in the legged-locomotion sense, denotes the zero-shot transfer of a learned policy πθ(a|s,m) across a distribution of robot morphologies, where $m \in M$ encodes static robot structure (limb lengths, mass, joint limits), and $s$ comprises all dynamic sensor observations (Bohlinger et al., 2 Sep 2025). After large-scale training with both fixed and extremely randomized morphologies, πθ can directly control previously unseen robots—including those with never-encountered inertial and actuation properties—without per-robot adaptation.

In the RL formalism, this extends to the Controlled Embodiment MDP ($CE$-MDP) (2405.14073):

$$\mathcal{M}_e^c = (\mathcal{S}_e, \mathcal{A}_e, \mathcal{P}_e, \gamma) \qquad \forall e \in \mathcal{E}$$

with global shared policy $\pi: \mathcal{S} \to \Delta(\mathcal{A})$ that must generalize over the union of $e$-indexed MDPs, via state or action projection as necessary.

Foundational to the theory is the hypothesis of "embodiment scaling laws": increasing the diversity of distinct morphologies in the training set leads to monotonic improvements in zero-shot transfer to unseen embodiments, more so than simply scaling the amount of data per embodiment (Ai et al., 9 May 2025).

In HCI, trans-embodiment generalizes further as the ability of an artificial agent to persist through “teleportation” (migration) across platforms or avatars, with identity coherence formalized via higher-order epistemic logic (e.g., beliefs about unique personal facts establishing agent sameness in the mind of the user) (Angel et al., 2018). In human-object-agent interaction, the concept extends to the asynchronous, imaginative mapping of objects and human-related identities onto artificial conversational agents (Xu et al., 18 Jan 2026).

2. Neural and Algorithmic Approaches for Robust Trans-Embodiment

Architecture-level Strategies

• Morphology-aware neural policies (URMAv2): These architectures encode each joint’s static properties (rotation axis, link length, mass, torque limits) and dynamic state via attention mechanisms, producing joint-aware latent representations $\bar z_{{\rm joints}}$ and action decoders that remain agnostic to the number and type of degrees of freedom (Bohlinger et al., 2 Sep 2025).
• Embodiment-aware Transformers (EAT): Sequence modeling policies condition future action prediction on a token stream that recurrently includes explicit embodiment descriptors, ensuring the autoregressive process adapts to each morphology’s constraints (Yu et al., 2022).
• SE(3)-Equivariant Policies: Actions are represented and predicted in end-effector frames, with an analytic decoder enforcing equivariance with respect to any transformation of base/end-effector frames, ensuring robust generalization across coordinate system changes (Chen et al., 18 Sep 2025).
• Latent Embodiment-Agnostic Spaces: Skills, actions, or object affordances are encoded in shared latent spaces (contrastively or via prototype clustering), with decoders or projection heads mapping latents to embodiment-specific actions at inference (Bauer et al., 17 Jun 2025, Aktas et al., 2024, Xu et al., 2023).

Data and Curriculum Strategies

• Extreme Embodiment Randomization (ER): Rapid, performance-tuned randomization of morphology, actuation, and sensory parameters forces a policy to operate in a near-continuous family of embodiments, building invariance instead of memorization (Bohlinger et al., 2 Sep 2025).
• Cross-embodiment unsupervised pre-training (PEAC): Pre-training with intrinsic rewards that maximize trajectory distinguishability across embodiment classes—through a learned embodiment discriminator—yields policies with strong embodiment-awareness and fast post-hoc adaptation (2405.14073).
• Synthetic Continued Pretraining (SCP): For language-action models, SCP constructs synthetic multi-robot (e.g. bimanual) action sequences from unimanual data, allowing an autoregressive model to self-align to new embodiment output tokenization before any fine-tuning (Li et al., 3 Nov 2025).

3. Skill Representation, Affordance, and Action Translation Across Morphologies

• Affordance Equivalence Latent Spaces: Object-action-effect tuples from multiple robots are projected into a shared affordance code $z$, with explicit equivalence clustering enforcing semantic alignment of skills across distinct agent morphologies. At test time, a demonstration by robot $R_1$ can be mapped—without correspondence labels—into valid command sequences for robot $R_2$ that produces the same effect (Aktas et al., 2024).
• Cross-embodiment Skill Prototypes: By temporal clustering of video encodings, skill segments from both human and robot videos are mapped onto shared prototype anchors; a conditioned diffusion policy translates these embeddings into action sequences for the target robot (Xu et al., 2023).
• Latent Action Unification via Contrastive Learning: Encoders translate high-dimensional, heterogeneous action spaces (e.g., human hand, five-fingered robot, two-fingered gripper) into a unified latent. Co-training and cross-decoding in this latent boost policy transfer and allow heterogeneous embodiment control with a single backbone (Bauer et al., 17 Jun 2025).
• Scene Flow and Object-Centric Planning: By predicting the 3D scene flow of manipulated object parts (rather than joint paths), embodiment-agnostic plans can be generated and subsequently inverted to suit any manipulator, human or robot (Tang et al., 2024).

4. Trans-Embodiment in Human-Object-Agent Interaction and Identity

• Formalized Agent Teleportation: Epistemic modal logics formally specify criteria under which a human user perceives an agent as the “same individual” when it moves between embodiments, focusing on shared beliefs about unique personal memories (Angel et al., 2018).
• Triadic Trans-Embodiment in CA Design: Identity work during life transitions can be scaffolded by a chatbot that shifts between imagined object, person, or user self-representations, supporting fluid, asynchronous trans-embodiment between human, object, and synthetic agent (Xu et al., 18 Jan 2026).
• Parallel Embodiment/Agency: Pilots (e.g., disabled café workers) can operate multiple avatars simultaneously, distributing their agency and expressive gestures across distinct robot bodies and tasks, facilitated by interface unification and shared gesture/command protocols (Barbareschi et al., 2023).

5. Empirical Findings and Scaling Law Evidence

• Locomotion: URMAv2 with ER achieves zero-shot transfer to novel quadrupeds (≈90% of nominal reward) and even to challenging humanoid robots (40–70%), without any fine-tuning (Bohlinger et al., 2 Sep 2025).
• Manipulation: Cross-embodiment methods such as XSkill and TrajSkill exhibit large increases in success rates for manipulation—from human video prompts only—compared to goal-conditioned or embedding-free baselines (up to 16.7% absolute improvement) (Tang et al., 9 Oct 2025, Xu et al., 2023).
• Scaling Laws: Adding distinct morphologies to the training set consistently improves generalization to unseen embodiments, with performance gains saturating as the number of per-embodiment data increases but continuing to rise with morphology diversity (Ai et al., 9 May 2025).
• Vision-Language-Action: Policies built for equivariance, or warmed-up via SCP and augmented with explicit task-embodiment graphs, achieve substantially higher real-world and simulation success rates in multi-robot, multi-arm tasks versus state-of-the-art VLA models (Li et al., 3 Nov 2025, Chen et al., 18 Sep 2025).

Table: Selected Policy and Representation Approaches for Trans-Embodiment

Approach	Principle	Key Result(s)
URMAv2 + ER (Bohlinger et al., 2 Sep 2025)	Morphology-attention, curriculum	90% zero-shot on quadrupeds, robust on real robots
EAT (Yu et al., 2022)	Embodiment-tokenized Transformer	Zero-shot real-world transfer, stable under morphology changes
PEAC (2405.14073)	Embodiment-informative intrinsic RL	Strong cross-embodiment adaptation in RL with no rewards
Affordance Equivalence (Aktas et al., 2024)	Shared latent clustering	Imitation from human push to robot action with 90% success
XSkill (Xu et al., 2023)	Skill prototypes	85%+ cross-embodiment manipulation, skill compositionality
SE(3)-equivariant PL (Chen et al., 18 Sep 2025)	SE(3) action-space equivariance	Few-shot transfer, robust under camera/base changes

6. Open Problems, Limitations, and Future Directions

• Representational bottlenecks: Most policies rely on URDF-style descriptors or low-dimensional morphology vectors; richer, possibly graph-based morphology embeddings are needed for more diverse robots (Bohlinger et al., 2 Sep 2025, Yu et al., 2022).
• Failure on out-of-distribution embodiments: Highly divergent topologies, inaccurate or absent descriptors, or radical sensor/actuator gaps can still defeat existing policies (Bohlinger et al., 2 Sep 2025, Ai et al., 9 May 2025).
• Generalization outside locomotion: Manipulation, whole-body tasks, and aerial or deformable robots present additional representational and dynamic challenges.
• Hierarchy and modularity: Multi-level (high-level skill/intent, low-level dynamics) and multi-time-scale architectures are necessary to scale from simple tokenized actions to rich, compliant, and adaptive strategies (Hoffmann et al., 15 May 2025).
• Identity and social continuity: For agent migration in user-facing applications, formal/logical approaches to preserving and certifying continuity of identity/personality remain idealized and need empirical integration (Angel et al., 2018).
• Scaling and co-design: Research into “embodiment scaling laws” is ongoing; it remains open how architectural and sampling choices (e.g., number and type of base robots) control generalization bounds (Ai et al., 9 May 2025).

7. Broader Implications and Impact

Trans-embodiment is a foundational principle for building general-purpose robotic, virtual, and social agents. By focusing on morphology-aware policies, shared affordance spaces, explicit equivariance, and robust curriculum/randomization, research has made tangible progress in realizing agents that can “become many bodies” with no per-robot adaptation. At the same time, challenges around representation, embodiment realism, scalability, and social continuity thread through both technical and human-centered applications. Advances in this field not only underpin the next generation of generalist robotic control and imitation but also drive theoretical and empirical understanding of transferable agency and distributed identity in embodied intelligence (Hoffmann et al., 15 May 2025, Bohlinger et al., 2 Sep 2025, Angel et al., 2018, Xu et al., 2023).