HumiTransfer turns two-person video into humanoid interaction data. It is still research.

The result

A new paper, HumiTransfer, describes a system for taking video of two people interacting and retargeting one person’s movement to a humanoid robot. The target is not a general-purpose companion robot. It is a research pipeline for interactions such as a handshake, a dance, or a hug, where the robot has to preserve distance, orientation, contact, and balance relative to another person.

The authors report processing 1,200 in-the-wild videos into a dataset they call HumiMimic, with about 500,000 frames for a 26-degree-of-freedom Unitree R1 Edu and a 23-degree-of-freedom HighTorque Mini-hi. The paper says the system was tested in real-robot deployments. RoboSignal’s research listing independently identifies the work as a research result, not a deployment or product announcement.

That distinction matters. A robot reproducing an interaction under a research setup is a narrower claim than a robot that can interpret a household member’s intent, choose appropriate contact, and recover safely when the person moves unexpectedly.

How it works

The system starts with ordinary video containing two people. It reconstructs their three-dimensional motion, including foot-ground contact, then maps one participant’s movement onto a target humanoid. Rather than copying joint positions alone, the method weighs relationships between the two bodies: where they face, how far apart they are, which body parts contact, and which joints matter most for a particular interaction.

That is the useful technical idea. A handshake fails if a robot merely waves an arm through a human-shaped pose. The hand has to arrive in the right place, at the right time, while the robot remains upright and avoids a collision. HumiTransfer adds a later optimization stage meant to maintain balance, contact, and collision avoidance after the initial retargeting.

The resulting dataset is then used to train motion-generation methods that follow a human partner. The authors compare their results with existing baselines and report better motion fidelity, preserved inter-body relationships, and contact accuracy across the two research platforms. Those are paper results, not an independently audited benchmark or a home trial.

What it means for a home robot

Home robots will need to work around people, not just around objects on a table. Socially aware movement could matter for handing over a glass, assisting an older person, moving through a crowded kitchen, or simply avoiding an awkward collision. Video is also abundant compared with carefully captured robot demonstrations, so a method that draws structured motion data from real-world recordings could reduce one data bottleneck.

But the paper’s examples point to a harder lesson: human interaction is not a single motion clip. A person can pull away, misread a robot’s approach, change pace, or make contact in an unexpected place. A household system would need perception, intent recognition, force limits, clear consent boundaries, and reliable recovery behavior alongside motion transfer. None of that follows automatically from an improved retargeting pipeline.

The hardware scope is narrow, too. The authors target the Unitree R1 Edu and HighTorque Mini-hi research humanoids. Results on those embodiments do not establish that other humanoids, much less a consumer robot, can perform the same interactions.

Caveats

This is a preprint submitted to arXiv on October 27, 2026. It has not been presented here as peer-reviewed evidence, and the public paper is the main source for its performance claims. The authors report physical deployments, but the paper does not turn those demonstrations into evidence of unsupervised household use.

The source material also concerns motion transfer, not a complete safety system. Contact-rich tasks require careful evaluation because getting the geometry right is only one part of interacting with a person. The paper’s collision-avoidance and balance stages are encouraging engineering details, but they are not a substitute for a consumer safety case.

For now, HumiTransfer is best read as a data and motion-generation contribution. It shows a plausible way to make a humanoid’s interaction training less dependent on bespoke motion capture. It does not show that a home humanoid is ready to give a reliable handshake, much less to handle the messy social judgment that comes with living among people.