ReContact teaches a Unitree G1 to use its body while moving an object. It is still research.

The result

A new paper from researchers at Tohoku University reports a way to turn ordinary human-object video into contact plans for a Unitree G1 humanoid. The point is not merely to make the robot copy an arm path. It is to tell the controller which parts of the robot should touch an object, where, and in which direction they should press while a transfer unfolds.

The authors call the system ReContact. Their most concrete hardware demonstration has a G1 pick up a box from the side, support it against its thigh, shift to a two-palm hold, walk while loaded, reverse the regrasp, and set the box down. The project page describes a 1.9-kilogram box in that sequence and shows tests with six boxes from 0.7 to 1.9 kilograms.

That is a useful capability to study. A person routinely uses a hip, thigh, forearm, or torso to stabilize something that is awkward to hold. Most robot demonstrations still treat the hands as the only legitimate contact point. ReContact asks what happens when the rest of the body becomes part of the plan.

How it works

The system begins with video of a person interacting with an object. It reconstructs the object motion and creates what the paper calls an object-centric contact plan: phases of an action, the robot and object surfaces expected to meet, contact anchors, directions of force, and whether a contact should stick or slide.

That matters because matching an object trajectory alone can hide a bad grasp. An object can reach its destination while the robot squeezes at the wrong point, loses support during a transition, or drops it later. ReContact adds rewards for making contact at the planned point and in the planned direction, then trains a residual policy on top of a whole-body motion tracker.

The paper reports two simulation comparisons worth separating from the hardware work. In a thigh-supported regrasp task, ReContact reached the planned two-palm bottom hold in 83.5% of episodes across three training seeds, versus 0.16% for the best adapted baseline. In another tube-transfer test, the authors generated 247 references for different placements; their family policy raised in-workspace success from 20.4% for a policy tied to the source reference to 97.8%. Those are controlled research results, not a measure of domestic reliability.

What it means for a home robot

A home robot will eventually need to manage items that do not behave like neat factory parts: a bulky parcel, a grocery bag, a bottle that must stay upright, or laundry that changes shape while it is handled. Planning support across the body could make some of those handoffs more robust than a hand-only approach.

The G1 demonstrations also make the work more informative than a simulation-only paper. The authors report seven contact-rich manipulation tasks and two loco-manipulation tasks on real hardware. The project page shows examples including a tissue-pack transfer, spray-bottle grasp, banana grasp, and box transfers. That is evidence that the pipeline can reach a real humanoid, rather than only produce attractive motion in a renderer.

But this is still a lab workflow. The paper gives the controller a defined task, known objects, video-derived references, and an evaluation setup. It does not show a robot deciding on its own how to unpack an unfamiliar box in a cluttered kitchen, sharing space safely with people, or recovering from an unexpected slip without a prepared contact plan.

Caveats

The central demonstrations are narrow and deliberately structured. The paper does not establish how often the G1 can repeat these skills over long periods, how it handles objects outside the tested shapes and masses, or whether the same approach works amid household clutter and changing light. The reported percentage gains come from simulation benchmarks, not a field trial in homes.

There is also a difference between generating a useful reference from a human video and acquiring enough suitable video to cover ordinary household work. A robot that can transfer one box with thigh support is not yet a robot that can safely put away groceries.

For now, ReContact is a credible research result with a real hardware demonstration. It shows one route toward whole-body manipulation. It does not change the status of the Unitree G1, and it is not evidence that a general-purpose home humanoid is ready for customers.