Retraining a humanoid robot's motion model from scratch every time you swap hardware is expensive. Any2Any is a proposed shortcut.
Researchers introduced Any2Any, a transfer framework designed to port whole-body tracking (WBT) models — the systems that let humanoids imitate human motion — from one robot platform to another without starting over. The method works in two steps: first, it aligns the skeletal geometry of the source and target robots so the original model's inputs and outputs still make sense on new hardware; then it applies lightweight parameter-efficient fine-tuning to the parts of the model most sensitive to physical dynamics. In tests, the approach transferred Sonic models trained on the Unitree G1 to two LimX platforms, Oli and Luna, using roughly 1% of the compute a full training run would require.
The cost problem here is real. Whole-body control models need large motion-capture datasets and significant GPU time to train, which means every new humanoid form factor is a fresh bill. If transfer learning can cut that to a rounding error, the practical effect is that smaller labs — or companies fielding multiple robot variants — no longer need to treat each hardware revision as a clean-slate training project. That changes the economics of humanoid deployment more than any single benchmark score.
The catch, as with most transfer learning papers, is that "competitive or superior" performance on selected test platforms doesn't guarantee the approach generalizes cleanly to bodies with significantly different kinematics. The humanoid arms race has no shortage of form factors, and the real test will be whether Any2Any holds up as platforms diverge further from the Unitree G1 it was built around.