A learned controller helped a human-guided humanoid shift a basketball and a box between its arms and torso while keeping the objects supported. In the researchers’ real-robot tests, it completed 45 of 50 trials across three tasks, compared with 9 of 50 for direct motion retargeting and 2 of 50 for a generalist whole-body controller.
Bulky objects need more than a handoff
A robot that can grasp an object with its hands may still struggle to move a bulky item from one hold to another. The object can lose support while the arms change position. People often solve this by bracing a box against the chest or sliding it along an arm, but a humanoid needs to coordinate those contacts as it moves.
The researchers call this ability whole-body dexterity. Their controller uses the robot’s arms and torso as contact surfaces, so an object can shift between holds rather than remain fixed in a two-handed grasp.
The controller refines a person’s motion commands
The team first generated a library of stable ways to hold objects against different body regions. It then trained reinforcement-learning specialists to move objects between those holds in simulation and distilled their behavior into one controller for a real Unitree G1 humanoid.
A human operator still steers the robot. A motion-capture suit turns the operator’s movements into coarse commands, while the learned controller adjusts the arm targets to keep the object supported as the robot changes its hold. The robot’s lower-body controller handles balance. This arrangement lets the policy refine a person’s intended motion; it does not give the robot an independent task plan.
One robot and two objects set the limit
In the physical tests, the G1 moved a basketball or a 15-centimeter cardboard box from a two-handed hold to its shoulder or under one arm and back. A separate test asked it to pick up the box from a shelf, tuck it under one arm, carry it while walking, then place it on a table. Across these three task types, the learned controller completed 45 of 50 trials. The direct-retargeting baseline completed 9, and the SONIC generalist controller completed 2.
The comparison is promising, but the test was narrow: one robot model, two objects and one trained operator. The work is an arXiv preprint, and the reported trials do not establish that the controller will handle other humanoids, heavier loads or less controlled environments. The useful result is that a learned policy could preserve support through arm-and-torso handoffs in these lab tasks while a person continued to guide the motion.
Sources
Figure: Mc Gartoll et al., “Whole-Body Dexterity,” arXiv:2610.22184v1, CC BY-SA 4.0.