Data Analogies Enable Efficient Cross-Embodiment Transfer

Jonathan Yang, Chelsea Finn, Dorsa Sadigh

TL;DR

Scaling robot datasets isn't just about "more data"—it's about the right structure. This paper demonstrates that while visual diversity helps robots handle new viewpoints, transferring skills across different robot bodies (morphologies) requires "data analogies": paired demonstrations where different robots perform the same task with aligned trajectories. By prioritizing these analogies, the authors improved real-world cross-embodiment transfer success by 22.5% over standard large-scale datasets.

What Problem Does This Paper Address?

The robotics community is currently in a "scaling" era, aggregating massive datasets like Open X-Embodiment (OXE) to train generalist Vision-Language-Action (VLA) models. While these models show impressive visual robustness, it remains unclear if they are truly learning universal physical principles or just benefiting from massive visual regularization.

Specifically, when a robot has a new arm or gripper (a morphology shift), simply adding more random data from other robots often fails to help. The authors identify a "connectivity" gap: we lack a principled understanding of how to organize heterogeneous data so that a demonstration from a Kinova arm actually helps a Franka arm. They seek to move beyond "unstructured diversity" toward a data-centric strategy for high-fidelity transfer.

Methods

The researchers conducted controlled experiments across three primary domain shifts: Viewpoint (camera pose), Morphology (gripper/arm kinematics), and Appearance (textures/lighting). They compared two orthogonal axes of data collection:

  1. Coverage Strategy:
    • Targeted: Selecting data that specifically fills gaps relative to the target robot.
  2. Diverse: Sampling broadly and randomly across all available setups.
  3. Cross-Robot Pairing:
    • Unpaired: Independent demonstrations.
  4. Task-Paired: Different robots performing the same task (same objects/goals) but with unaligned movements.
  5. Trajectory-Paired (Data Analogies): Using Dynamic Time Warping (DTW) to align the execution strategies of different robots. This ensures that at any given timestep, both robots are in a similar stage of the task (e.g., both are just about to grasp the object).

They used a $\pi_{0.5}$-style VLA architecture, fine-tuning a pretrained model on a mixture of target-robot data and "translation" data (the analogies).

Datasets

  • Simulation: A benchmark based on RoboCasa, featuring tasks like "Pick and Place Counter to Sink" and "Flip Mug Upright." They used three robot embodiments (Kinova, Kinova3, UR5e).
  • Real-World: Experiments conducted on Franka Emika Panda, WidowX-250, and PiperX platforms.
  • Baselines: The models were compared against training on OXE (Open X-Embodiment) and Bridge+DROID datasets.
  • Augmentation: They used DALL-E 3 to perform "robot inpainting," synthetically changing the appearance of the robot in existing datasets to increase visual diversity without new collections.

Key Insights and Ideas

  • Visual vs. Physical Generalization: The "Aha!" moment is that different shifts require different data. Viewpoint and appearance shifts are primarily perceptual; they benefit most from Diverse coverage because it regularizes the vision encoder. However, morphology shifts are physical; they benefit most from Trajectory-Paired data.
  • The Morphology Plateau: Without pairing, increasing the diversity of robot arms barely improves performance. The model sees different control distributions (actions) but can't reconcile them with the visual scene. Analogies act as a "Rosetta Stone," teaching the model how different "bodies" map to the same "intent."
  • Composition Over Volume: The paper argues that the composition of the dataset is more important than raw scale. A smaller, carefully paired dataset can outperform a massive, unpaired one (like OXE) because it provides the structural "glue" needed for transfer.

Key Results

  • Simulation: The "Data Analogy" approach achieved an average of 19% higher success rate than standard training on large open-source datasets. In complex tasks like "Flip Mug," the gains were as high as 35-40 percentage points.
  • Real-World: On tasks like "Pen in Cup" and "Book on Bookshelf," the proposed method improved success by an average of 22.5% over OXE-based baselines.
  • Zero-Shot Failure: Policies trained only on the source data (e.g., BRIDGE) often achieved 0% success in new environments, while the analogy-tuned models remained robust.

Limitations and Future Work

  • Fixed Budget Constraints: The study focused on a few-shot regime (50 demonstrations). It is unclear if the advantage of analogies diminishes as the target-robot dataset grows to thousands of demonstrations.
  • Architecture Specificity: The experiments were tied to a specific VLA backbone ($\pi_{0.5}$). Future work could explore if larger models (like RT-2) or different action heads (e.g., diffusion vs. flow-matching) change these data requirements.
  • Sim-to-Real Gap: While the trends held in the real world, the "translation" data still relies on some manual staging to ensure "identical scenes" for pairing. Automating the discovery of analogies in unstructured, "in-the-wild" data remains a major open challenge.

Summary updated

Abstract

Generalist robot policies are trained on demonstrations collected across a wide variety of robots, scenes, and viewpoints. Yet it remains unclear how to best organize and scale such heterogeneous data so that it genuinely improves performance in a given target setting. In this work, we ask: what form of demonstration data is most useful for enabling transfer across robot set-ups? We conduct controlled experiments that vary end-effector morphology, robot platform appearance, and camera perspective, and compare the effects of simply scaling the number of demonstrations against systematically broadening the diversity in different ways. Our simulated experiments show that while perceptual shifts such as viewpoint benefit most from broad diversity, morphology shifts benefit far less from unstructured diversity and instead see the largest gains from data analogies, i.e. paired demonstrations that align scenes, tasks, and/or trajectories across different embodiments. Informed by the simulation results, we improve real-world cross-embodiment transfer success by an average of $22.5\%$ over large-scale, unpaired datasets by changing only the composition of the data.
BibTeX
@misc{yang2026data,
  author = {Jonathan Yang and Chelsea Finn and Dorsa Sadigh},
  title = {{Data Analogies Enable Efficient Cross-Embodiment Transfer}},
  year = {2026},
  eprint = {2603.06450},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2603.06450},
}
Figure