Weave, a new robotics framework, teaches humanoid robots to grab and carry objects using full-body coordination copied from human movement, not just scripted hand motions.
Researchers describe the system in a paper posted to arXiv (arXiv:2609.16683, https://arxiv.org/abs/2609.16683) on September 16, 2026. Weave converts recorded human-object interactions into robot-executable motion through what the team calls contact-aware retargeting and approach-motion completion. A single policy then commands 29 body joints and 12 finger joints at once, coordinating balance, walking, and grip. Across nine test objects, the system hit a 92.5% success rate on interactions it was trained on and 65.0% on sequences it had never seen. The team also released roughly 9,000 physically executed robot rollouts, about 23 hours of contact-annotated trajectory data, on the project site at https://xiaohu-art.github.io/Weave/.
Humanoid robots have gotten decent at walking and standing up after a shove. Coordinated object handling, using legs, torso, and fingers together while the object and the robot both move, has lagged further behind. Weave's approach sidesteps hand-coded contact planning by learning it from human demonstration video instead, which is cheaper to collect than scripted trajectories for every object and grip.
The bigger tell is the gap between the two numbers: 92.5% on rehearsed objects versus 65.0% on new ones. That drop is the actual research problem, not the demo reel, and it is the number to watch as labs push this kind of imitation learning past nine curated objects and into anything resembling a real room.