A new motion dataset gives humanoids more ways to hold onto objects. It is not a chore skill.

A humanoid needs more than a video of a person opening a cabinet or carrying a box. It needs a version of that motion that fits its joints, respects its feet, and stays in contact with the object. A paper posted to arXiv on September 28 tackles that translation problem. HOI-Retarget is a research method and dataset, not a robot product or a demonstrated home assistant.

The authors say they have released code and retargeted motion data. Their accompanying dataset page identifies the work with the paper and lists Unitree G1 and H2 among its humanoid targets. That makes it a potentially useful ingredient for researchers training physical skills. It does not show that either robot can independently complete a household task.

The result

The paper focuses on human-object interaction, or HOI: motions where a body and an object have to stay coordinated. Rather than treating a recorded human movement as a shape to copy, the method treats labeled contacts as targets in the object's frame. In plain English, the system tries to preserve where hands or other body parts meet the thing being handled while adapting the rest of the movement to a robot.

The authors say the process balances body tracking, foot support and smoothness under the target robot's kinematic limits. They also say one demonstration can be augmented across object sizes, and that contacts reconstructed from monocular video can be used. Those are claims from the paper, not independent performance tests.

The public dataset viewer shows 13.9k rows in its default subset. Its fields include robot and object labels, frame counts, contact-link names, contact sequences, quality-control flags and motion data. The page also shows subsets named corolehoi, imhd2, neuraldome, omomo and parahome. A dataset record is evidence that the release exists; it is not evidence of real-world task success.

How it works

Human demonstrations rarely fit a humanoid directly. A robot has different limb lengths, joint limits, balance constraints and grippers. The paper uses windowed trajectory optimization to make the motion compatible with those constraints while targeting contacts relative to the object. This matters because a hand motion that looks close enough in a video may still miss a mug handle, knock over a bottle, or leave the robot unable to keep its balance.

The dataset page's contact and quality-control fields point to the kind of bookkeeping this work requires. A useful motion library needs to say which robot and object are involved, where contact occurred, and whether a sequence passed checks. That is more concrete than a collection of unlabeled demo clips.

What it means for a home robot

A household robot will need many examples of object handling before it can be dependable around cupboards, dishes, clothing and clutter. A release that maps demonstrations onto named humanoid bodies could help researchers produce more of those examples than motion capture alone. The relevance is indirect but real: better training data can help make manipulation policies less brittle.

The missing parts are the parts a homeowner notices. The paper does not establish that a robot can recognize a new object, choose a safe grasp, recover after a slip, navigate a crowded kitchen, or finish a task repeatedly without supervision. It does not report a consumer deployment, a price, or a product timeline.

Caveats

HOI-Retarget is an arXiv preprint submitted to ICRA, not a peer-reviewed product evaluation. Its strongest claims concern data generation and motion transfer. We did not find an independent replication or a public benchmark comparing household task completion against other training pipelines.

The dataset is also released under a CC BY-NC-SA 4.0 license, according to its page. That may shape how companies can use it. For now, read this as a useful research release: it gives humanoid researchers more structured ways to translate human-object demonstrations, while leaving autonomous home help a long way downstream.