Whole-Body Dexterity: Humanoids That Brace and Carry with Arms and Torso

When a person picks up a big box, the hands are never the whole story — arms, chest and even thighs brace against it to share the load. Most robots only use their hands.

Background

Humanoid dexterous manipulation has advanced quickly, but public results remain largely confined to end-effectors and simple grasps, far short of the human ability to brace and slide bulky objects against the whole body. When an object exceeds what a single hand can hold stably, an end-effector-only approach struggles — and that is exactly the common case in warehousing, manufacturing and household carrying.

Core Innovation

Whole-Body Dexterity (WBD) is a general framework that treats the robot’s body and appendages as contact surfaces for grasping and manipulating objects. Its sim-to-real pipeline has three stages: first, it automatically generates a dataset of feasible whole-body grasps; next, it samples initial and goal grasps from that dataset and trains reinforcement-learning experts on an any-to-any repositioning objective; finally, it distills those experts into a single policy that uses only proprioception and object geometry as input. Because the input deliberately omits vision and object identity, the distilled policy generalizes across objects. At deployment the controller works interactively with a human teleoperator, who supplies only coarse motion references while the policy handles how to brace and adjust contact.

Results

• The distilled policy transfers zero-shot to a real Unitree G1, repositioning a basketball and a cardboard box to regions around the torso.
• The robot completes a long-horizon real-world task: stowing a box under one arm, carrying it from a shelf to a table, and placing it.
• On whole-body manipulation tasks, the distilled controller completes 90% of trials, well above 18% for direct retargeting and 4% for a generalist whole-body controller.

Limitations

The abstract reports real-world results for only two objects, a small sample, and gives no quantified robustness envelope across object size, weight and friction. The controller still relies on a human teleoperator for coarse references, so autonomy is partial, and the abstract does not address force-control safety during contact or whether sustained bracing risks overheating or wear. All of this needs validation at larger hardware scale.

Industry Implications

Whole-body manipulation matters because it directly raises the size and weight ceiling of what a robot can handle. In logistics unloading, manufacturing feeding and household moving, robots mostly face objects beyond a single-hand grip, and WBD offers a viable route to “carry it with the whole body.” Notably, the distilled policy uses only proprioception and object geometry, meaning deployment needs no expensive visual calibration and hardware cost stays bounded. The gap between 90% and 18%/4% shows whole-body contact is not a nice-to-have trick but the dividing line between success and failure. For teams building around humanoids, it suggests the next frontier of dexterous manipulation may be the body, not the fingers.