Data Analogies Enable Efficient Cross-Embodiment Transfer

Jonathan Yang; Chelsea Finn; Dorsa Sadigh

cs.ROMar 6, 2026

Data Analogies Enable Efficient Cross-Embodiment Transfer

Authors: Jonathan Yang, Chelsea Finn, Dorsa Sadigh

Organizations: Stanford University

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.

Figures & tables

Fig. 1: Cross-Embodiment Data Analogies. We study how to collect data so that demonstrations from one robot directly help another. Our data-centric recipe composes breadth ( coverage across viewpoints, morphologies, and scenes). Then, we search different data scaling strategies to find one that leads to high performance under a fixed budget. We find that datasets with high pairing between scenes and tasks as well as high coverage lead to high transfer performance.

Fig. 2: Domain Shift Axes. We study the role of data diversity and pairing across three domain shift axes: end-effector morphology, camera perspective, and visual appearance.

Fig. 3: Coverage versus Pairing. Simulation images depicting the data collection strategies. Coverage is the diversity of data on the generalization axis, while pairing is the similarity of the tasks or trajectories in the data.

Fig. 4: Sim and Real-World Robots. We train a cross-embodiment policy to transfer the tasks of putting a pen in a cup and putting a book on a bookshelf to a new robot. We evaluate with the Franka Emika Panda, WidowX, and ARX Piper robot arms.

Fig. 5: Main coverage plot (Success Rate). Success Rate (%) on the target robot across Coverage $\times$ Pairing for (a) Viewpoint, (b) Morphology, (c) Appearance. Error bars: 95% CI. Dashed lines show Target-only (few-shot) and Target upper bound (same extra budget on target).

Fig. 6: Comparison with large open-source datasets. Success Rate (%) on four RoboCasa tasks when training on (i) narrow two-robot data ( Bridge+DROID ), (ii) large unpaired open-source datasets ( OXE ), and (iii) our OXE+Translational composition that reweights coverage and adds trajectory-level alignment. Results are shown for two target robots ( Panda , Jaco ); error bars denote 95% CI.

Fig. 7: Effect of scaling diversity. Success Rate (%) on the target robot as we increase the diversity of source data (distinct embodiments, viewpoints, and scenes). Curves compare Naïve/Uniform , Targeted Coverage , and Trajectory-Paired . Dashed lines show the 50-shot Target-only baseline and the Target upper bound .

Fig. 8: Real-robot transfer. Success Rate (%) on two tabletop tasks under three embodiment settings: PiperX $\!\to$ WidowX, WidowX $\!\to$ Franka, and WidowX $\!\to$ Piper. Bars compare a few-shot target baseline, a narrow two-robot pool ( Bridge+DROID ), a diversity-weighted open-source pool ( OXE ), and our composed OXE+Translational . Error bars denote 95% CI.

Fig. 9: BRIDGE dataset transfer. We evaluate how well our method leads to transfer on two BRIDGE tasks across two environments each.

Fig. 10: Simulation Robots. Simulation robots with diverse camera angles and objects on a countertop scene.

Fig. 11: Real Robots. Real robots with diverse camera angles and objects on a kitchen scene for the pen in cup task.

Dataset	Split
Bridge	$12.5\%$
DROID	$12.5\%$
Fractal	$12.5\%$
Taco Play	$12.5\%$
Jaco Play	$12.5\%$
Roboturk	$12.5\%$

TABLE I: Data Splits

Hyperparameter	Value
Base model	Pi0.5 VLA
Fine-tuning strategy	LoRA
Batch size (global)	32
Optimizer	AdamW
Learning rate schedule	Cosine decay with warmup
Peak learning rate	$5\times 10^{-5}$

TABLE II: Pi0.5 LoRA Training Hyperparameters. Typical values used across our experiments.

Domain	Coverage	Unpaired	Task-Paired	Trajectory-Paired
Viewpoint	Targeted	$45.0\pm 3.0$	$50.0\pm 3.0$	$52.0\pm 3.0$
Viewpoint	Diverse	$64.0\pm 3.0$	$68.0\pm 3.0$	$70.0\pm 3.0$
Morphology	Targeted	$24.0\pm 4.0$	$46.0\pm 4.0$	$62.0\pm 4.0$
Morphology	Diverse	$28.0\pm 4.0$	$48.0\pm 4.0$	$64.0\pm 4.0$
Appearance	Targeted	$48.0\pm 2.5$	$55.0\pm 2.5$	$57.0\pm 2.5$
Appearance	Diverse	$54.0\pm 2.5$	$62.0\pm 2.5$	$68.0\pm 2.5$

TABLE III: Success rates (%) across coverage and pairing strategies for different domain shifts. Results show mean ± 95% CI over 100 simulation trials.

Task	Robot	Baseline	Bridge+DROID	OXE	OXE+Translational
PnP Counter→Sink	Panda	$18.0\pm 2.5$	$18.0\pm 2.6$	$23.0\pm 3.0$	$\mathbf{40.0\pm 3.1}$
PnP Counter→Sink	Jaco	$12.0\pm 2.5$	$18.0\pm 2.6$	$22.0\pm 3.0$	$\mathbf{38.0\pm 3.1}$
PnP Sink→Counter	Panda	$16.0\pm 2.5$	$24.0\pm 2.6$	$27.0\pm 3.0$	$\mathbf{39.0\pm 3.1}$
PnP Sink→Counter	Jaco	$10.0\pm 2.5$	$24.0\pm 2.6$	$26.0\pm 3.0$	$\mathbf{36.0\pm 3.1}$
Flip Mug	Panda	$22.0\pm 2.5$	$28.0\pm 2.6$	$33.0\pm 3.0$	$\mathbf{68.0\pm 3.1}$
Flip Mug	Jaco	$20.0\pm 2.5$	$28.0\pm 2.6$	$32.0\pm 3.0$	$\mathbf{62.0\pm 3.1}$

TABLE IV: Success rates (%) comparing our compositional approach with large-scale open-source training. Results for Panda and Jaco target robots across four simulation tasks.

Task	Transfer	Baseline	Bridge+DROID	OXE	OXE+Translational
Pen in Cup	PiperX→WidowX	$50\pm 3.5$	$60\pm 3.0$	$65\pm 3.0$	$\mathbf{85\pm 3.0}$
	WidowX→Franka	$40\pm 4.0$	$50\pm 3.5$	$50\pm 3.0$	$\mathbf{75\pm 3.0}$
	WidowX→Piper	$55\pm 3.5$	$65\pm 3.5$	$65\pm 3.0$	$\mathbf{90\pm 3.0}$
Book on Bookshelf	PiperX→WidowX	$25\pm 2.5$	$35\pm 3.0$	$40\pm 3.0$	$\mathbf{65\pm 3.0}$
	WidowX→Franka	$20\pm 3.0$	$30\pm 3.0$	$35\pm 3.0$	$\mathbf{60\pm 3.0}$
	WidowX→Piper	$30\pm 2.5$	$40\pm 3.0$	$45\pm 3.5$	$\mathbf{65\pm 3.5}$

TABLE V: Real-world success rates (%) across different robot transfer pairs. Results averaged over 5 seeds with 95% confidence intervals.

Domain	Method	Baseline	2	5	10	20	40
Viewpoint	Naive (Uniform)	35.0	$45.0$	$48.0$	$52.0$	$58.0$	$60.0$
	Targeted Coverage		$50.0$	$57.0$	$61.0$	$66.0$	$68.0$
	Trajectory-Paired		$52.0$	$61.0$	$66.0$	$71.0$	$74.0$
Morphology	Naive (Uniform)	24.0	$24.0$	$25.5$	$28.0$	—	—
	Targeted Coverage		$46.0$	$46.8$	$48.0$	—	—
	Trajectory-Paired		$62.0$	$62.8$	$64.0$	—	—

TABLE VI: Performance scaling with source diversity. Shows success rates at different numbers of source embodiments/viewpoints/scenes for each method.

Explore similar work

Latent Policy Steering: An Efficient and Flexible Framework for Cross-Embodiment Transfer
Sep 18, 2026Yiqi Wang, Mrinal Verghese, Jeff SchneiderCross-Embodiment Transfer Visuomotor Policy Learning
Robotics Institute, School of Computer Science, Carnegie Mellon University, 5000 Forbes Ave, Pittsburgh, United States
Cross-Embodiment Transfer via Behavior-Aligned Representations
Jul 30, 2026Ajay Sridhar, Jensen Gao, Jonathan Yang +3Representation Learning Vision-Language-Action Models
1Stanford University · 2Toyota Research Institute
Intelligence Across Embodiments
Sep 22, 2026Bo Ai, Henrik I. Christensen, Hao SuCross-Embodiment Robot Learning Embodied AI
Stanford University · University of California San Diego · Sudo AI GmbH

Sep 18, 2026J/K move · Enter open · S save

Latent Policy Steering: An Efficient and Flexible Framework for Cross-Embodiment Transfer

Yiqi Wang, Mrinal Verghese, Jeff Schneider

Robotics Institute, School of Computer Science, Carnegie Mellon University, 5000 Forbes Ave, Pittsburgh, United States

Cross-Embodiment Transfer Visuomotor Policy Learning Robot Policy Generalization Data-Efficient Robot Learning World Models for Robotics

Open paper AskPDF arXiv

The performance of learned robot visuomotor policies depends heavily on the size and quality of their training data, yet collecting high-quality demonstrations remains costly for robots in the real world. Although large-scale robot and human datasets are increasingly available, embodiment gaps and mismatched action spaces make them difficult to leverage directly. Cross-embodiment transfer, reusing experience from other embodiments to improve learning on a target embodiment, is therefore crucial for scaling robot learning beyond per-robot data collection. In this work, we find that efficient transfer can be achieved by learning from what is shared across embodiments, the visual dynamics of how the world responds to motion, and by effectively exploiting the scarce target-embodiment data at test time. The proposed framework, called Latent Policy Steering (LPS), implements an embodiment-agnostic pretraining phase, which trains an image-based World Model (WM) with optical flow across diverse embodiments. The resulting WM is finetuned on the target embodiment with robot actions. It then steers the base policy toward better actions by searching in the WM's latent space for plans that stay close to the finetuning data. LPS is a policy-agnostic framework: it can flexibly accommodate different policies without having to retrain them. In Robomimic and real-world evaluations, LPS improves the average performance of Diffusion Policy relatively by 16% and 62%, and Pi0.5 by 8% and 14%, with only 50 demonstrations on an unseen target embodiment.

Abstract

Explore similar work

Latent Policy Steering: An Efficient and Flexible Framework for Cross-Embodiment Transfer

Cross-Embodiment Transfer via Behavior-Aligned Representations

Intelligence Across Embodiments