Cross-Embodiment Data Curriculum

• Cross-embodiment data curriculum is a framework that structures heterogeneous demonstration data by balancing coverage, pairing, abstraction, and supervision to bridge differences in morphology, kinematics, and appearance.
• It employs staged organization and progressive abstraction—using methods like dense-to-sparse motion transfer and quality filtering—to enhance transferability across differing embodiments.
• Empirical studies show that explicit pairing and structured data composition yield marked improvements in real-world tasks, with gains of up to 22.5% under specific configurations.

Searching arXiv for papers on cross-embodiment learning and data curriculum to ground the article in the current literature. Cross-embodiment data curriculum denotes the deliberate organization, abstraction, staging, and weighting of data used to transfer manipulation competence across embodiments whose morphology, kinematics, action spaces, viewpoints, or visual appearance differ. In the recent literature, the term does not refer to a single algorithmic template. Instead, it encompasses several related strategies: balancing coverage across embodiment-relevant axes, constructing paired demonstrations or “data analogies,” progressively abstracting embodiment-specific observations into invariant motion representations, ordering supervision regimes from rich to sparse or from synthetic to real, and filtering or weighting heterogeneous data according to transferability or quality. Across these formulations, the common objective is to expose a policy or generative model to data that preserves task semantics while reducing embodiment-specific confounds, thereby turning heterogeneous demonstrations into a bridge rather than an obstacle for transfer (Yang et al., 6 Mar 2026).

1. Conceptual basis and problem formulation

A central premise of cross-embodiment learning is that heterogeneous demonstrations are only useful if their variation is organized along the dimensions that matter for transfer. One explicit formulation models a dataset as

$\mathcal{D}=\{(e,\tau)\},$

where $e\in\mathcal{E}$ indexes embodiment and $\tau=\big((o_1^e,a_1^e),\ldots,(o_T^e,a_T^e)\big)$ is a trajectory. In this setting, the target robot $e^\star$ may have only a few demonstrations $\mathcal{D}_{e^\star}^{\text{few}}$, and source embodiments are valuable only insofar as they provide a bridging signal during fine-tuning or representation learning (Yang et al., 6 Mar 2026).

The literature distinguishes several axes of cross-embodiment variation. One line of work isolates end-effector morphology, robot platform appearance, and camera perspective, emphasizing that morphology is action-dependent, whereas appearance and viewpoint are primarily perceptual (Yang et al., 6 Mar 2026). Other work frames the gap more broadly as differences in morphology, kinematics, degrees of freedom, and visual appearance between, for example, human bodies and robot manipulators (Tang et al., 9 Oct 2025). In dexterous settings, the problem is often stated as the lack of a shared state-action representation across hands with different numbers of degrees of freedom and different contact geometries (He et al., 3 Nov 2025).

Within this landscape, “curriculum” has at least two technical meanings. In a narrow sense, it denotes a staged progression in training signals, as in dense-to-sparse motion supervision or synthetic continued pretraining followed by embodiment-specific fine-tuning. In a broader data-centric sense, it denotes structured dataset composition: balancing morphology exposure, pairing demonstrations across robots, or grouping mixed-quality trajectories by ordinal quality. Several papers are explicit that they are not classical curriculum learning methods even though they contain curriculum-like mechanisms, such as view augmentation with decaying dropout or training-time data augmentation pipelines (Rayyan et al., 23 Sep 2025).

2. Principal curriculum dimensions

The most explicit decomposition of cross-embodiment data curriculum is the distinction between coverage and pairing. Coverage concerns which portions of the embodiment-relevant space are represented: targeted coverage fills gaps relative to the target robot, whereas diverse coverage samples broadly across available robots, scenes, or views. Pairing concerns correspondence structure: unpaired data contain no explicit alignment, task-paired data share the same task instance and objects, and trajectory-paired data align execution strategies across embodiments. The empirical result is that these dimensions are orthogonal and contribute differently depending on the shift axis (Yang et al., 6 Mar 2026).

A second curriculum dimension is abstraction level. TrajSkill converts raw human videos into dense optical flow, then into sparse optical flow trajectories, then into robot manipulation videos, and finally into executable actions. This progression is presented as both an explicit dense-to-sparse staged training regime and an implicit hierarchy of data abstraction that reduces embodiment dependence at each stage (Tang et al., 9 Oct 2025). A related abstraction strategy appears in cross-embodiment world models, where hands and objects are represented as sets of 3D particles and actions as particle displacements, thereby replacing embodiment-specific joint spaces with a shared geometric space (He et al., 3 Nov 2025).

A third dimension is supervision type. OmniHumanoid separates unpaired embodiment-specific videos from paired motion-aligned cross-embodiment data. Stage I trains embodiment LoRAs $\Phi_e$ on unpaired videos $\mathcal{U}_e$, while Stage II freezes the LoRAs and trains the shared motion transfer model on paired data $\mathcal{D}_{\mathrm{pair}}$. Adaptation to a new embodiment then requires only a new LoRA trained on unpaired videos $\mathcal{U}_{e^\ast}$ (Song et al., 12 May 2026). ET-VLA similarly uses a two-stage sequence: Synthetic Continued Pretraining on synthetic multi-robot token sequences, followed by fine-tuning on real target-embodiment data (Li et al., 3 Nov 2025).

A fourth dimension is data quality. In cross-embodiment inverse reinforcement learning from mixed-quality demonstrations, XIRL-Buckets groups trajectories into ordinal quality buckets and applies temporal cycle-consistency only within each bucket. This is explicitly a quality-based curriculum rather than an embodiment-ordering curriculum (Mattson et al., 2024). X-Diffusion addresses an analogous issue at the action level: human demonstrations supervise diffusion training only at noise levels where the noised human action becomes indistinguishable from robot actions (Pace et al., 6 Nov 2025).

3. Explicit curriculum mechanisms

Several papers instantiate explicit, stage-wise curricula rather than merely structured data composition.

TrajSkill provides a clear dense-to-sparse curriculum. From a human video, it computes dense optical flow $\mathbf{F}_t \in \mathbb{R}^{H \times W \times 2}$ with RAFT, samples candidate points on a grid, draws keypoints with probability proportional to initial flow magnitude,

$e\in\mathcal{E}$0

propagates them by flow integration,

$e\in\mathcal{E}$1

and smooths the sparse flow field with Gaussian filtering. The downstream generator is trained first with Dense Flow Supervision and then with Sparse Trajectory Alignment so that dense motion teaches accurate dynamics and sparse trajectories match the embodiment-invariant test-time prompt. The resulting conditional generative-control formulation is

$e\in\mathcal{E}$2

This is one of the clearest examples of a cross-embodiment data curriculum by representational abstraction and staged supervision (Tang et al., 9 Oct 2025).

X-Diffusion implements a curriculum over diffusion noise levels. A classifier $e\in\mathcal{E}$3 predicts whether a noised action comes from a robot or a human. For each human action sequence, the method defines a trajectory-specific threshold

$e\in\mathcal{E}$4

Robot data supervise all noise levels, whereas human data contribute only when $e\in\mathcal{E}$5. The policy objective therefore includes a hard inclusion rule rather than a soft weighting: $e\in\mathcal{E}$6 The curriculum is not over tasks or embodiments directly; it is over the degree to which embodiment-specific execution details have been washed out by the forward diffusion process (Pace et al., 6 Nov 2025).

ET-VLA contributes a different staged curriculum. Standard OpenVLA-style control expects $e\in\mathcal{E}$7 action tokens, but bimanual control requires $e\in\mathcal{E}$8. Synthetic Continued Pretraining constructs synthetic multi-robot supervision by cross-sampling within a batch $e\in\mathcal{E}$9: for a sample $\tau=\big((o_1^e,a_1^e),\ldots,(o_T^e,a_T^e)\big)$0, a different sample $\tau=\big((o_1^e,a_1^e),\ldots,(o_T^e,a_T^e)\big)$1 is chosen and the predicted action tokens $\tau=\big((o_1^e,a_1^e),\ldots,(o_T^e,a_T^e)\big)$2 are appended to $\tau=\big((o_1^e,a_1^e),\ldots,(o_T^e,a_T^e)\big)$3, yielding a synthetic $\tau=\big((o_1^e,a_1^e),\ldots,(o_T^e,a_T^e)\big)$4-token target. This stage warms up the autoregressive model to the new embodiment’s token grammar before real-target fine-tuning (Li et al., 3 Nov 2025).

OmniHumanoid likewise uses a two-stage curriculum, but its organizing principle is supervision type rather than token structure: first learn appearance and morphology from unpaired videos, then learn transferable motion from paired motion-aligned data while freezing embodiment adapters and isolating motion conditioning from embodiment-specific modulation (Song et al., 12 May 2026).

4. Data interfaces and embodiment-invariant representations

Cross-embodiment data curricula often depend on a common interface in which demonstrations from distinct bodies become comparable.

TrajSkill’s interface is a 2D sparse trajectory language derived from optical flow. The paper argues that the transferable content of a demonstration is motion intent rather than the source body itself: a hand moving a cup, a robot gripper moving a block, or a sphere agent moving through space can all be expressed as sparse optical flow trajectories. This suggests that curriculum design can proceed by stripping away appearance and morphology while preserving spatiotemporal dynamics (Tang et al., 9 Oct 2025).

CEI constructs a unified interface through functional similarity. A source demonstration

$\tau=\big((o_1^e,a_1^e),\ldots,(o_T^e,a_T^e)\big)$5

is rewritten as a target demonstration $\tau=\big((o_1^e,a_1^e),\ldots,(o_T^e,a_T^e)\big)$6 for a new embodiment $\tau=\big((o_1^e,a_1^e),\ldots,(o_T^e,a_T^e)\big)$7. Embodiments are represented by point-direction pairs $\tau=\big((o_1^e,a_1^e),\ldots,(o_T^e,a_T^e)\big)$8, and similarity is quantified with Directional Chamfer Distance: $\tau=\big((o_1^e,a_1^e),\ldots,(o_T^e,a_T^e)\big)$9 Trajectory alignment then solves

$e^\star$0

with warm-started sequential optimization $e^\star$1. In this formulation, the data curriculum is realized through demonstration re-authoring: one source demonstration seeds executable target demonstrations across embodiments (Wu et al., 14 Jan 2026).

Cross-embodiment world models propose a different interface based on embodiment-invariant dynamics. For embodiment $e^\star$2, the hand is represented as particles

$e^\star$3

objects as

$e^\star$4

and actions as particle displacements

$e^\star$5

A graph-based world model predicts

$e^\star$6

Here the curriculum is not a staged scheduler but a scaling law over training embodiments: broader embodiment diversity improves generalization to unseen hands, particularly when simulation and real human data are co-trained in a balanced mixture (He et al., 3 Nov 2025).

MV-UMI is instructive as a boundary case. It is explicitly characterized as a cross-embodiment multi-view interface and data-preprocessing framework rather than a classic curriculum. Its contribution is to combine egocentric and third-person views, remove the human in the third-person view by SAM-2 segmentation and inpainting, and apply a decaying random view-dropout schedule $e^\star$7. This is best interpreted as implicit curriculum via view augmentation and decaying corruption, not as a formal curriculum learning pipeline (Rayyan et al., 23 Sep 2025).

5. Empirical findings on what data organization helps

A recurrent empirical conclusion is that more data and better-composed data are not equivalent. Controlled experiments on morphology, appearance, and viewpoint shifts show that perceptual shifts benefit from broad diversity, whereas morphology shifts benefit far more from pairing than from unstructured diversity. For viewpoint, performance rises from 45% under Targeted + Unpaired to 64% under Diverse + Unpaired, and to 70% under Trajectory-paired + Diverse. For appearance, the corresponding values are 48%, 54%, and 68%. For morphology, however, the key jump is from 24% under Targeted + Unpaired and 28% under Diverse + Unpaired to 46% under Targeted + Task-paired, 62% under Targeted + Trajectory-paired, and 64% under Diverse + Trajectory-paired. In real-world transfer, changing only the composition of the data improves success by an average of 22.5% over large-scale, unpaired datasets (Yang et al., 6 Mar 2026).

The same principle appears in retargeting and generation pipelines. CEI transfers demonstrations from a Franka Panda source robot to 16 embodiments across 3 tasks in simulation and supports bidirectional transfer between a UR5+AG95 and a UR5+Xhand across 6 real-world tasks, achieving an average transfer ratio of 82.4%. The real-world average success across the six tasks is reported as 70% (Wu et al., 14 Jan 2026). TrajSkill shows that motion-abstraction curricula can improve both generative fidelity and execution: on MetaWorld it achieves FVD 318.83 and KVD 268.03, versus CogVideo’s 528.04 and 422.86, summarized as 39.6% lower FVD and 36.6% lower KVD; for execution on MetaWorld it attains 44.7% overall success rate, and in a Franka Panda kitchen setup for “Put the Banana in the Basket” it reaches 90.9% pick success and 81.8% place success (Tang et al., 9 Oct 2025).

Noise-aware and supervision-aware curricula also yield measurable gains. Across five real-world manipulation tasks, X-Diffusion achieves a 16% higher average success rate than the best baseline (Pace et al., 6 Nov 2025). ET-VLA raises real-world average success from 5/77 for OpenVLA to 46/77 for the full system, that is, from about 6.49% to about 59.74%, and the paper summarizes the gain over OpenVLA on six real-world tasks as 53.2% (Li et al., 3 Nov 2025). OmniHumanoid, while centered on cross-embodiment video generation rather than policy learning, shows that paired motion learning plus unpaired adaptation can enable scalable transfer to unseen humanoid embodiments without retraining the shared motion model (Song et al., 12 May 2026).

The broader data-augmentation literature further supports the data-centric view. RoVi-Aug is not a classical curriculum, but by augmenting demonstrations across robots and viewpoints it rebalances the training distribution toward the cross-product of robots and views; in physical experiments, co-training on original and augmented data improves success rates by up to 30% (Chen et al., 2024). X-DiffVLA similarly does not define a formal curriculum schedule, yet its embodied forcing and morphology-aware denoising structure produce 64.5% average success on RoboCasa and 71.0% on Isaac Gym, with reported improvements of 15.3% and 12.5% over state-of-the-art baselines (Li et al., 24 May 2026).

Approach	Curriculum principle	Representative result
TrajSkill	Dense flow to sparse trajectories	39.6% lower FVD, 36.6% lower KVD on MetaWorld (Tang et al., 9 Oct 2025)
Data Analogies	Coverage plus pairing	22.5% average real-world gain over large unpaired datasets (Yang et al., 6 Mar 2026)
X-Diffusion	Noise-level gating of human supervision	16% higher average success rate than the best baseline (Pace et al., 6 Nov 2025)
CEI	Functional retargeting into target embodiment data	82.4% average transfer ratio (Wu et al., 14 Jan 2026)
ET-VLA	Synthetic continued pretraining then fine-tuning	46/77 vs 5/77 real-world average success (Li et al., 3 Nov 2025)

These results support a general interpretation: morphology transfer requires explicit structure, whereas many perceptual shifts can be handled by sufficiently broad coverage.

6. Misconceptions, limitations, and open directions

A common misconception is that any cross-embodiment method using heterogeneous data is a data curriculum. The literature is more precise. MV-UMI is primarily a multi-view interface and domain-alignment method, not a formal curriculum algorithm (Rayyan et al., 23 Sep 2025). RoVi-Aug is a training-time augmentation pipeline rather than an annealed easy-to-hard schedule (Chen et al., 2024). Conversely, TrajSkill, X-Diffusion, ET-VLA, and OmniHumanoid each contain explicit stage-wise ordering mechanisms, but they order supervision, abstraction, or token structure rather than task difficulty in the classical sense (Tang et al., 9 Oct 2025).

A second misconception is that broad, unpaired diversity is universally sufficient. Controlled evidence contradicts this for morphology. Viewpoint and appearance benefit from diverse coverage, but morphology transfer saturates under unstructured diversity and improves most when demonstrations are paired at the task or trajectory level (Yang et al., 6 Mar 2026). This suggests that a cross-embodiment curriculum must be matched to the dominant shift axis rather than applied uniformly.

A third limitation concerns data and representation quality. TrajSkill depends on accurate optical flow and may struggle when flow estimation is noisy, occluded, or subtle (Tang et al., 9 Oct 2025). MV-UMI depends on segmentation quality, inpainting, and the usefulness of the third-person camera under occlusion (Rayyan et al., 23 Sep 2025). XIRL-like cross-embodiment reward learning fails on mixed-quality, mixed-embodiment demonstrations because temporal alignment assumptions break when not all trajectories are near-optimal, motivating human-feedback-based reward learning or quality-aware bucketing (Mattson et al., 2024).

Benchmark evidence also indicates that cross-embodiment curricula remain far from solved for truly out-of-distribution morphology generalization. AnyBody distinguishes interpolation, composition, and extrapolation regimes and finds that interpolation is easiest, composition harder, and extrapolation hardest; in some extrapolation settings, multi-embodiment agents achieve 0% success. This suggests that balanced multi-embodiment training alone does not induce robust structural abstraction, and that future curricula may need explicit progression by morphology distance, structural novelty, or informativeness-weighted sampling (Parakh et al., 21 May 2025).

A plausible implication is that future cross-embodiment data curricula will combine several of the mechanisms that are currently studied separately: paired analogies for morphology, broad coverage for perceptual robustness, abstraction into embodiment-invariant interfaces, quality-aware filtering, and staged supervision from synthetic, dense, or strongly aligned data toward sparse, unpaired, or deployment-realistic data. The recent literature collectively indicates that cross-embodiment transfer improves when data are not merely accumulated, but composed so that correspondence, abstraction, and coverage are made explicit (Yang et al., 6 Mar 2026).