Scaling H2R — Scaling Laws and Diversity in Human-to-Robot

About the Workshop

From whether to how much

This workshop continues the Human-to-Robot (H2R) line of inquiry on learning robot skills from human data, but deliberately narrows the question. Where prior efforts established whether robots can be taught from sensorized, modeled human behavior, the central premise here is that the binding constraint has now shifted from feasibility to scale and diversity: human data is a source of embodied experience that is abundant and cheap to collect, yet we still lack an empirical understanding of how robot capability grows as a function of how much human data we gather, how diverse it is, and how far the human embodiment sits from the target robot.

We bring together researchers in imitation learning, egocentric perception, dexterous manipulation, hardware co-design, and world modeling to ask what it actually takes to turn large, heterogeneous human data into reliable robot policies.

Fig 1. The workshop's central question — do returns keep accumulating along hours, demonstrators, scenes & embodiments, or saturate? And where does the dashed projection actually go?

Data engines

Researchers building human data engines and retargeting pipelines.

Scaling laws

Those studying scaling laws and dataset composition.

World models

Those developing egocentric world models from human video.

Hardware co-design

Those co-designing hardware that closes the human–robot gap.

Core Challenges

Research Questions

Scaling laws for human data

Does robot policy performance improve predictably with the volume of human data, and along which axes — hours, tasks, demonstrators, scenes, embodiments — do returns actually accumulate versus saturate?

Diversity over quantity

Which forms of diversity (task, scene, object, viewpoint, demonstrator morphology) drive downstream generalization, and how do we measure and compose for diversity rather than merely inflate dataset size?

The embodiment gap at scale

How do morphological, kinematic, and contact/force mismatches degrade transfer as data scales, and can hardware co-design converge the two embodiments to make scaling effective?

Egocentric world models

Can action-conditioned, predictive world models learned from large-scale egocentric human video yield representations or simulators that transfer to robot control?

Data engines and evaluation

What pipelines for collection, retargeting, automatic labeling, and quality filtering scale to in-the-wild human data — and how do we benchmark whether added human data genuinely improves robot policies?

Program

Tentative Schedule

08:30 – 08:35Opening Remarks

08:35 – 09:05Oral Session 1Spotlights 1–4 · 5 min each + 2 min Q&A

09:05 – 09:30Invited Talk 1

09:30 – 09:355 min break

09:35 – 10:05Oral Session 2Spotlights 5–8 · 5 min each + 2 min Q&A

10:05 – 10:30Invited Talk 2

10:30 – 11:00Coffee Break & Poster Session

11:00 – 11:25Invited Talk 3

11:25 – 11:50Invited Talk 4

11:50 – 12:00Open ForumCrowdsourced audience questions via QR code

12:00 – 12:25Panel Discussion

12:25 – 12:30Awards & Closing Remarks

Call for Papers

Contribute your work

We solicit contributed papers through a single Regular Papers Track. We invite new, preliminary, and in-progress research as well as position papers on scaling and diversifying human data for robot learning.