Cross-Embodiment Normalizer

• Cross-Embodiment Normalizer is a mechanism that maps heterogeneous observations and actions into a unified, embodiment-agnostic interface.
• It separates shared task intent from embodiment-specific execution, using methods like canonical slotting, geometric alignment, and latent factorization.
• Empirical results show improved few-shot transfer and zero-shot generalization, boosting performance across varying robotic and human agents.

Searching arXiv for papers on cross-embodiment normalization and related methods. A Cross-Embodiment Normalizer is a representation, interface, or architectural mechanism that reduces embodiment-specific variation while preserving task-relevant structure, so that policies, world models, planners, or generative systems can transfer across robots, humans, or other agents with different kinematics, morphologies, sensing, and action parameterizations. In the recent literature, the term does not denote a single standardized module; rather, it appears as a recurring design pattern implemented through canonicalized action motifs, geometry-aware latent spaces, semantic joint interfaces, particle-based state-action abstractions, embodied intent compressors, tool-centric motion invariants, and related constructs. Across these formulations, the common function is to separate a shared notion of what should happen from embodiment-specific details of how it is realized (Zhi et al., 14 Feb 2026).

1. Definition and scope

In cross-embodiment learning, the central difficulty is that heterogeneous embodiments induce mismatched state and action spaces. One robot may output end-effector velocities, another joint torques, another a dexterous hand configuration, while human demonstrations may be given only as video or motion trajectories. A Cross-Embodiment Normalizer addresses this mismatch by mapping embodiment-specific observations, actions, trajectories, or latent features into a shared interface that is more comparable across embodiments (Zhi et al., 14 Feb 2026).

Several papers make this role explicit. MOTIF describes action motifs as a “normalizing intermediate representation” that decouples embodiment-agnostic spatiotemporal patterns from robot-specific execution (Zhi et al., 14 Feb 2026). CEI frames a unified interface that transfers demonstrations by preserving “functionally equivalent contact and interaction behaviors” through directional-aware geometric alignment (Wu et al., 14 Jan 2026). Tenma defines a cross-embodiment normalizer as a data-unification mechanism that standardizes heterogeneous state and action spaces into fixed-length, masked, normalized vectors mapped to a shared latent space (Davies et al., 15 Sep 2025). OPFA places the normalizer in a geometry-aware latent action space that supports a shared decoder across grippers and hands without embodiment-specific decoder tuning (Mu et al., 15 Mar 2026).

The same abstraction also appears outside direct manipulation policy transfer. XIRL learns an embodiment-invariant task-progress embedding from videos and uses distance to a goal embedding as a reward for reinforcement learning (Zakka et al., 2021). XMoP normalizes over manipulators by representing an embodiment as a fixed-layout sequence of whole-body link poses in $SE(3)$, combined with masking and frame augmentation (Rath et al., 2024). In cross-lingual representation learning, Iterative Normalization enforces zero-mean and unit-norm constraints to make embedding spaces more alignable, and the paper explicitly notes that the same principles can generalize to “Cross-Embodiment” alignment beyond languages (Zhang et al., 2019). This suggests that the term covers both robotics-specific interfaces and broader normalization strategies for heterogeneous representation spaces.

2. Core design principle: separating shared structure from embodiment-specific execution

A recurring formulation is to factor cross-embodiment transfer into a shared component and an embodiment-specific component. MOTIF states this directly: motifs encode “what to do,” while embodiment-specific state/action encoders and decoders determine “how to execute” in a target action space $A_e^i$ (Zhi et al., 14 Feb 2026). In OPFA, the shared component is the Geometry-Aware Latent Representation $z \in \mathbb{R}^k$, while embodiment-specificity is realized only by a fixed selection matrix $S_e$ that extracts the relevant joints from a universal decoder output $\Theta \in \mathbb{R}^D$ (Mu et al., 15 Mar 2026). In BLM$_1$, a Perceiver-based intent-bridging interface compresses MLLM hidden states into embodiment-neutral guidance tokens $z_t$, after which a shared DiT is wrapped by embodiment-specific state encoders, action encoders, and action decoders (Tan et al., 28 Oct 2025).

This factorization can be implemented at different levels of abstraction. Some methods normalize actions. MOTIF learns discrete codes over canonicalized end-effector trajectory segments using vector quantization, progress-aware alignment, and embodiment adversarial invariance (Zhi et al., 14 Feb 2026). OPFA learns a latent action space from reachable-state point clouds derived from forward kinematics (Mu et al., 15 Mar 2026). Tenma canonicalizes joint and controller signals into fixed slot layouts with per-embodiment min–max scaling and masks (Davies et al., 15 Sep 2025). Other methods normalize states or dynamics. The particle-based world-model framework in “Scaling Cross-Embodiment World Models for Dexterous Manipulation” maps different embodiments to sets of 3D particles and defines actions as particle displacements, so that the learned dynamics model operates in an embodiment-invariant space (He et al., 3 Nov 2025). XMoP uses a common whole-body $SE(3)$ pose representation over a fixed token layout with morphology masking (Rath et al., 2024).

A different branch normalizes task semantics or intent rather than control signals. XIRL uses temporal cycle-consistency to learn an embedding that captures shared task progress across videos with different embodiments (Zakka et al., 2021). “Bridging the Embodiment Gap” explicitly disentangles task semantics $z_{\text{task}}$ from embodiment factors $z_{\text{emb}}$ by minimizing mutual information and applying dual InfoNCE regularization (Li et al., 5 May 2026). OmniHumanoid factorizes shared motion transfer from embodiment-specific rendering adapters, with a unidirectional attention mask that prevents embodiment-specific priors from leaking into the motion-conditioning branch (Song et al., 12 May 2026). In these cases, the normalizer is less an action-space converter than a mechanism for preserving embodiment-invariant content.

A plausible implication is that Cross-Embodiment Normalizers can be categorized by the level at which invariance is imposed: trajectory-level, action-level, state-level, dynamics-level, intent-level, or representation-geometry level. The surveyed papers suggest that the most effective designs make this factorization explicit rather than relying on an oversized shared backbone to discover it implicitly.

3. Main technical realizations

The literature contains several distinct realizations of the normalizer concept.

Realization	Representative paper	Mechanism
Discrete spatiotemporal motifs	MOTIF (Zhi et al., 14 Feb 2026)	VQ-VAE on canonicalized end-effector segments with progress-aware alignment and embodiment adversarial constraints
Functional geometry interface	CEI (Wu et al., 14 Jan 2026)	Directional Chamfer Distance over end-effector contact points and normals
Adaptive feature modulation	AdaMorph (Zhang et al., 12 Jan 2026)	AdaLN conditioned by pooled robot prompts
Soft prompt conditioning	X-VLA (Zheng et al., 11 Oct 2025)	Embodiment-specific prompt tokens injected into a shared Transformer
Fixed canonical slot standardization	Tenma (Davies et al., 15 Sep 2025)	Masked, normalized fixed-length state/action vectors
Tool-centric motion invariants	LEGATO (Seo et al., 2024)	DHB motion-invariant descriptors over handheld gripper trajectories
Whole-body pose canonicalization	XMoP (Rath et al., 2024)	Fixed-layout link $A_e^i$0 tokens with morphology masking and frame augmentation
Particle state-action abstraction	World models for dexterous manipulation (He et al., 3 Nov 2025)	Particle displacements and graph dynamics
Intent compression	BLM$A_e^i$1 (Tan et al., 28 Oct 2025)	Perceiver compression of frozen MLLM hidden states into fixed intent tokens

In MOTIF, canonicalization is applied to end-effector trajectory segments by translating and rotating the raw path into a local frame anchored at $A_e^i$2 and normalizing by workspace scale. A VQ-VAE then learns motif codes $A_e^i$3 with commitment loss, while a phase-weighted InfoNCE objective and an adversarial embodiment classifier push the motifs toward temporal consistency and embodiment invariance (Zhi et al., 14 Feb 2026). This is a particularly explicit instance of a Cross-Embodiment Normalizer because the codebook itself is intended to be shared across robots.

CEI implements normalization geometrically. It represents candidate contact sites as point-normal pairs and aligns source and target embodiments by minimizing the Directional Chamfer Distance,

$A_e^i$4

with $A_e^i$5 in experiments (Wu et al., 14 Jan 2026). The normalizer here is not a latent vector but a shared notion of functional similarity.

AdaMorph normalizes inside the network by conditioning layer normalization statistics on embodiment prompts. Its decoder applies

$A_e^i$6

so the same morphology-agnostic latent intent can be projected onto different robot-specific manifolds through layer-wise modulation (Zhang et al., 12 Jan 2026). X-VLA uses a related but architecturally simpler strategy: separate learnable soft prompts per data source are concatenated with fused multimodal tokens, steering the shared Transformer’s attention landscape while adding only 0.04% non-shared parameters in X-VLA-0.9B (Zheng et al., 11 Oct 2025).

Tenma and XHugWBC emphasize semantic alignment through canonical indexing. Tenma uses fixed canonical slots plus per-embodiment masks and min–max scaling into $A_e^i$7 before shared MLP projection (Davies et al., 15 Sep 2025). XHugWBC defines a global joint space of fixed dimension $A_e^i$8 with a fixed joint ordering, zero-padding, and a controllability mask $A_e^i$9, so that both observations and actions live in semantically aligned canonical coordinates across humanoids (Xue et al., 5 Feb 2026). This suggests that even relatively simple slot-based or joint-index-based interfaces can function as Cross-Embodiment Normalizers when combined with structure-aware architectures.

4. Canonicalization, invariance, and alignment objectives

Most Cross-Embodiment Normalizers rely on explicit canonicalization before learning. In MOTIF, trajectory canonicalization consists of translation-rotation anchoring and workspace scale normalization (Zhi et al., 14 Feb 2026). LEGATO transforms tool-frame motion into Denavit-Hartenberg Bidirectional motion invariants, using relative increments and directional-change descriptors rather than raw $z \in \mathbb{R}^k$0 signals (Seo et al., 2024). XMoP expresses all link poses in the robot base frame, converts them to a continuous 9D pose representation, and applies fixed token layout, morphology masking, and frame augmentation to make the representation robust to link-frame convention changes (Rath et al., 2024). The particle world-model framework defines the shared state as end-effector and object particles in a common world frame, with actions as end-effector particle displacements (He et al., 3 Nov 2025).

Alignment losses vary substantially. MOTIF uses a soft-weighted InfoNCE objective over averaged segment embeddings with phase weights

$z \in \mathbb{R}^k$1

so that segments with the same instruction and similar task phase are pulled together (Zhi et al., 14 Feb 2026). Its adversarial objective

$z \in \mathbb{R}^k$2

is applied through gradient reversal to suppress embodiment cues in latent tokens (Zhi et al., 14 Feb 2026). “Bridging the Embodiment Gap” instead minimizes an upper bound on mutual information between $z \in \mathbb{R}^k$3 and $z \in \mathbb{R}^k$4 using CLUB, while separately enforcing intra-space consistency with InfoNCE (Li et al., 5 May 2026). XIRL replaces contrastive matching with temporal cycle-consistency over videos, using soft nearest-neighbor matching across demonstration sequences to arrange frames along a shared progress axis (Zakka et al., 2021).

Some papers normalize by architectural isolation rather than explicit penalties. OmniHumanoid uses branch-isolated attention with $z \in \mathbb{R}^k$5 and $z \in \mathbb{R}^k$6, and confines embodiment-specific LoRA adapters to the denoising branch (Song et al., 12 May 2026). BLM$z \in \mathbb{R}^k$7 freezes the MLLM during policy training and compresses hidden states into a fixed number of intent tokens with a Perceiver, thereby stabilizing the intent distribution across embodiments without an explicit alignment loss (Tan et al., 28 Oct 2025). This suggests that invariance can arise from information-flow restrictions, not only from auxiliary objectives.

The older Iterative Normalization work is notable because it makes the geometry of normalization fully explicit: repeated centering and $z \in \mathbb{R}^k$8 normalization enforce zero-mean and unit-norm constraints on embeddings,

$z \in \mathbb{R}^k$9

making the spaces more amenable to orthogonal alignment (Zhang et al., 2019). Although developed for cross-lingual word embeddings, the paper explicitly notes that the same normalization principles apply beyond languages. This suggests that some cross-embodiment problems may benefit from basic geometric conditioning even before more task-specific mechanisms are introduced.

5. Empirical evidence across robotics and embodied generation

The empirical literature reports that explicit normalization often improves few-shot transfer, zero-shot generalization, or cross-domain robustness. MOTIF reports that averaged transfer success improves by +6.5% over strong flow-matching baselines in simulation, and that in real-world settings it achieves 67.50% Transfer and 74.38% Global at 5-shot, yielding a +43.7% few-shot improvement over strong baselines (Zhi et al., 14 Feb 2026). Its ablations tie these gains directly to the normalizer: removing motif guidance, canonicalization, progress-aware alignment, or adversarial invariance reduces transfer (Zhi et al., 14 Feb 2026).

CEI demonstrates cross-embodiment transfer from a Franka Panda to 16 different embodiments across 3 simulated tasks and reports an average transfer ratio of 82.4% in real-world bidirectional transfer between UR5+AG95 and UR5+Xhand across 6 tasks (Wu et al., 14 Jan 2026). The ablation “without Direction” averages 32% success and fails grasp tasks, indicating that the directional component of the functional representation is part of the normalizing mechanism rather than an incidental feature (Wu et al., 14 Jan 2026).

OPFA reports that cross-embodiment co-training can improve success rates by more than 50% compared to single-source training, and that adding only eight demonstrations from a new embodiment can achieve performance comparable to that of a well-trained model with 72 demonstrations (Mu et al., 15 Mar 2026). Tenma reports an average success rate of 88.95% in-distribution, 72.56% under object shift, and 81.13% under scene shift, substantially exceeding baselines under matched compute (Davies et al., 15 Sep 2025). EmbodiSteer, while operating as an inference-time normalizer rather than a learned representation, reduces collision rate by 46.1% and improves task success rate by 28.5% across 9 simulated robots, with a 90.0% collision reduction and 36.7% success increase on two physical robots (Wang et al., 11 Jun 2026).

Beyond direct robot control, normalizer-style mechanisms also improve cross-embodiment video generation and motion retargeting. OmniHumanoid reports that without motion-appearance decoupling, Embodiment drops from 8.43 to 2.53 and Motion from 9.06 to 6.35, directly supporting the importance of isolation between transferable motion and embodiment-specific rendering (Song et al., 12 May 2026). “Bridging the Embodiment Gap” reports better cross-embodiment editing metrics than VACE and Phantom, and its ablation without the dual contrastive objective degrades fidelity while producing entangled embeddings (Li et al., 5 May 2026). AdaMorph reports strong zero-shot generalization across 12 humanoid robots, with median Pearson correlation coefficients above 0.8 for root velocity consistency and above 0.85 for whole-body activity consistency across embodiments (Zhang et al., 12 Jan 2026).

A broader pattern emerges across these results. Normalizers appear to be most beneficial when data are scarce, embodiments are highly heterogeneous, or transfer must occur with minimal retuning. This suggests that explicit separation of shared structure from embodiment-specific execution can act as a strong sample-efficiency prior.

6. Relation to shared-private architectures, misconceptions, and open directions

A common misconception is that a shared backbone with small embodiment-specific heads already constitutes sufficient cross-embodiment normalization. Multiple papers argue otherwise. MOTIF explicitly contrasts itself with shared-private architectures such as HPT and GR00T N1, stating that those methods rely on implicit alignment inside a shared trunk and often suffer from limited private capacity and lack explicit adaptation (Zhi et al., 14 Feb 2026). OPFA makes a related point by showing that naive co-training with separate decoders can overfit in few-shot settings and discard transferable geometric structure (Mu et al., 15 Mar 2026). X-VLA similarly argues that domain-specific output heads alone handle output heterogeneity but do not resolve cross-embodiment perception and reasoning shifts (Zheng et al., 11 Oct 2025).

Another misconception is that cross-embodiment normalization necessarily means a single explicit normalization layer. The surveyed papers show the opposite. In some cases the normalizer is a discrete codebook (Zhi et al., 14 Feb 2026); in others, a directional geometric metric (Wu et al., 14 Jan 2026), a latent factorization (Li et al., 5 May 2026), a set of soft prompts (Zheng et al., 11 Oct 2025), a particle representation (He et al., 3 Nov 2025), or a training-free joint-space steering mechanism (Wang et al., 11 Jun 2026). OmniHumanoid explicitly notes that it does not introduce a separate FiLM- or AdaIN-style module; instead, normalization is realized architecturally through branch isolation and adapter placement (Song et al., 12 May 2026). Thus, “Cross-Embodiment Normalizer” is best understood as a functional role rather than a fixed architectural primitive.

Several future directions recur across papers. MOTIF proposes hierarchical motifs, hybrid discrete-continuous codes, curriculum on phase alignment, and richer language integration (Zhi et al., 14 Feb 2026). CEI suggests embodiment-conditional decoders and multimodal motion generation extensions (Wu et al., 14 Jan 2026). OPFA points to stronger embodiment embeddings and physics-informed decoder constraints (Mu et al., 15 Mar 2026). BLM$S_e$0 indicates that richer embodiment-specific wrappers and calibration may be needed for extreme kinematic differences (Tan et al., 28 Oct 2025). XHugWBC points toward richer morphology embeddings, better contact modeling, and adaptive online normalization (Xue et al., 5 Feb 2026). A plausible implication is that future Cross-Embodiment Normalizers will increasingly combine multiple forms of invariance—geometry, phase, intent, semantics, and physical feasibility—rather than relying on a single interface.

Taken together, the literature presents Cross-Embodiment Normalization as a unifying principle for scalable transfer across heterogeneous agents. Whether implemented through motifs, geometry-aware latents, semantic joint layouts, embodied intent compression, particle abstractions, or disentangled latent variables, the central aim remains consistent: impose a canonical structure in which reusable task-relevant regularities are preserved and embodiment-specific variation is confined to a smaller, more manageable layer of adaptation (Zhi et al., 14 Feb 2026).