Researchers from Amazon, the University of California, San Diego, and the University of Michigan tested a Unitree G1 humanoid on tasks that required moving objects against its arms and torso. A skilled human teleoperator wearing a motion-capture suit guided every real-world trial, while the controller adjusted the robot’s arm commands to keep each object supported, according to the research report.
The controller completed 45 of 50 trials, a 90% success rate. That is the central result of Whole-Body Dexterity, or WBD, a method for using a humanoid’s arms and torso as contact surfaces, alongside its hands. The research team reports that WBD outperformed direct motion retargeting, at 18%, and a general-purpose whole-body controller, at 4%, in the tested tasks. The results appear in the WBD paper.
WBD moved a basketball and box across the robot’s body
The demonstrations tested a specific kind of manipulation: repositioning an object while changing where the robot supports it. The G1 moved a basketball and a 15-centimeter cardboard box to its shoulder or under one arm, then repositioned them. In a longer sequence, it took a box from a shelf, tucked it under an arm, walked to a table, and set it down.
That final sequence combines carrying with locomotion. The robot has to preserve contact as it walks, rather than complete a pick and place from a fixed position. It demonstrates a control strategy for a coupled task, not independent warehouse operation. The human operator’s role is central to that distinction: the reported 90% measures performance with a skilled person guiding the real-world trials, not an autonomous robot managing the task alone.
Simulation trained one controller to shift support
WBD treats the robot’s body and appendages as contact surfaces for grasping and moving objects. Instead of keeping an object in one hand-centered grasp, the robot can shift support across its arms and torso as the task changes. A box tucked under an arm, for example, can remain supported while the robot carries it.
The team generated about 2.5 million simulated grasps using the arms and torso. Reinforcement-learning experts learned to move objects between those grasps; the researchers then distilled the experts into a single controller. The controller uses the robot’s joint readings and rough object dimensions, according to the paper’s method description.
The team reports that this policy transferred zero-shot from simulation to a real Unitree G1. In this context, zero-shot means the policy was transferred without additional task-specific real-robot training reported by the authors. It does not remove the human operator from the real-world tests. The controller’s role was to adjust arm commands to maintain object support while the teleoperator guided the robot.
The 90% result points to a research bet, not a deployment case
The industrial implication is narrower than a claim that humanoids are ready for unstructured logistics. WBD shows that a controller built around whole-body contact can outperform the two reported baselines on a small set of real-robot tasks. It gives robotics teams a reason to test whether torso and arm support can make bulky-object handling more reliable when a hand-centered grasp is difficult.
The potential economic link is plausible but unproven. If a robot can support an object against its body, some tasks may rely less on a specialized gripper. But the trials report neither operating cost nor throughput, and they do not compare WBD with task-specific industrial systems. Those measurements are needed before the approach can be judged against existing automation on cost or reliability.
My forecast is that, within 24 months, leading industrial robotics teams will increase research spending on whole-body control relative to some single-task gripper work. The reasoning is conditional: WBD’s performance gap gives researchers a concrete control approach to investigate; simulation-generated grasps and a distilled policy offer a route to broadening the tested behaviors; and, if that approach works without continuous human guidance, it could apply to more varied handling tasks. None of those steps has yet established an industrial return.
The forecast would be weakened if follow-up trials cannot show reliable task completion without continuous teleoperation, or if whole-body control proves more costly or less dependable than task-specific grippers for the relevant jobs. The decisive evidence will be repeatable performance across useful workflows, with autonomy, speed, and cost measured alongside success rate. A 90% result across 50 guided trials is a research signal, not a deployment business case.
Teleoperation remains the next test for Whole-Body Dexterity
For physical automation, WBD’s result matters because it broadens the set of contact strategies tested on a humanoid. The G1 carried and repositioned a box using its arms and torso, relying on a controller trained through simulated grasps. That is a concrete demonstration of whole-body manipulation under the reported conditions, not evidence that robots can now handle unpredictable sites without supervision.
The next test is whether the same control approach can perform useful handling workflows with less human guidance and across a wider range of objects and conditions. Until those tests succeed, the motion-capture operator remains part of the system.
