A robotics research paper proposes a humanoid controller that works with whatever body tracking data you have on hand, not a complete skeleton.
The system, called AnyBody, trains a single motion model that can accept any subset of body keypoints — a few joints, a full skeleton, or anything in between — and still produce coordinated whole-body movement. Prior approaches either demanded expensive full-body motion capture or split upper- and lower-body control into separate systems, which tends to break down when a robot needs to move its legs and arms together. AnyBody addresses this with a teacher-student training pipeline: a privileged "teacher" model trained on a large motion corpus gets distilled into a leaner student model, and a transformer encoder with masked self-attention handles whatever keypoints arrive at inference time. A lightweight residual layer then handles task-specific fine-tuning without retraining the core model.
The practical upshot is that humanoid robots could be steered from partial or variable sensor data — useful for teleoperation scenarios where not every joint is tracked, or for deploying across different hardware setups. The researchers demonstrate locomotion, obstacle reaching, and in-air writing, which together suggest the system handles both coarse movement and fine motor tasks.
Most humanoid control research is still fighting the data bottleneck — full-body motion capture is slow and expensive to produce at scale. AnyBody's masked keypoint approach borrows a page from how masked language models handle missing tokens, and if it holds up outside the lab, it could make scalable robot policy learning considerably less painful.