Whole-Body Dexterity
This preprint from Amazon and university co-authors teaches a Unitree G1 humanoid to hold and shift bulky objects with its arms and torso, the way a person tucks a box under one arm, instead of relying on its hands alone. On a real robot the learned controller finished 90% of whole-body tasks, against 18% for direct motion retargeting and 4% for NVIDIA's SONIC whole-body controller.
Can a humanoid learn to grasp and reposition objects using any part of its body as a contact surface, including arm handoffs around the torso, and transfer that skill zero-shot from simulation to a real robot steered by a human operator?
Humanoid manipulation research mostly treats the hands as the only place where contact happens. Bulky objects such as boxes and balls are easier to carry when braced against the chest or held under one arm, but a two-armed robot has few independent supports, so releasing one arm means the other arm and the torso must keep the object pinned. Teleoperators cannot feel whether an object is slipping, which in practice limits teleoperation to hand-only pick-and-place.
Earlier whole-body work held bulky objects in a sustained hug in front of the torso, relied on tactile skins, or used model-based contact planning under planar or box-shaped assumptions. General whole-body motion trackers such as SONIC and retargeting tools such as GMR copy a human's motion onto the robot but have no notion of keeping an object supported while arms change grip.
The authors borrow the 'finger gaiting' idea from dexterous hands and apply it to the whole humanoid. They first generate 2.51 million simulation-stable whole-body grasps for 15 primitive objects (80 to 240 mm) by optimising contact on arm and torso links and filtering under four times gravity and sweeping disturbance wrenches. Reinforcement-learning experts (WBD-EX), one per object, learn to move an object between grasp modes (left arm, both arms, right arm) and body regions; every handoff must pass through a two-arm hold. The experts' 274,000 trajectories (873 million frames) are distilled into a single diffusion-based controller that uses proprioception and a five-number object shape descriptor, and refines a human operator's coarse motion commands into stable contact.
In simulation, with noise added to all 14 arm commands, the distilled controller increased mean time to failure by 70% on training objects and by 158% and 71% on a held-out sphere and cube. On a real Unitree G1 driven by an operator in an Xsens suit, it completed 45 of 50 trials (90%) across shoulder-tap and underarm-stow round trips with a size-3 basketball and a 15 cm box plus a shelf pick, stow, carry and place sequence, against 9 of 50 (18%) for GMR retargeting and 2 of 50 (4%) for SONIC. Underarm stowing of the box succeeded 10 of 10 times with the new controller, against 1 of 10 for GMR and 0 of 10 for SONIC. In the full shelf task it placed the box 7 times out of 10.
It is a preprint. All real-world trials used one skilled operator with more than eight hours of practice, so the result shows what the controller adds to a human, not autonomous behaviour. Objects were light primitives no larger than 24 cm; leg contact was excluded and the controller was trained without a lower-body controller. The robot still dropped objects in a few trials, and the method needs measured object dimensions. Each new object class needs its own expert, which the authors reduce by warm-starting from sphere experts.
Warehouse and logistics humanoids from Amazon's own fulfilment network to Agility, Figure and Boston Dynamics spend much of their time moving totes and boxes, and Amazon's authorship signals interest in using the torso as a load path rather than buying ever-stronger hands. For teleoperation data companies, a controller that keeps objects pinned while the operator moves could raise the yield of usable demonstrations. Expect this kind of whole-body grasp layer to appear inside humanoid stacks within one to two years, first under teleoperation.