Create and explore imaginative 3D worlds with our new product Dreamfield (opens in a new tab)

Generalization to heterogeneous humanoid skeletons

Generalize part-aware feature learning from the Unitree G1’s 23-DoF representation to heterogeneous humanoid skeletons.

Background

PartiGen is evaluated primarily on the Unitree G1, whose 23-DoF input layout and fixed joint-angle masks are specific to that morphology. Applying the representation to other humanoid platforms would require adapting these definitions and assessing the resulting model. The paper explicitly identifies generalization to heterogeneous skeletons as unresolved, making this an open problem for extending the method beyond the G1 platform.

References

The amount of retraining needed for such adaptation has not been established, and generalizing part-aware feature learning to heterogeneous skeletons remains an open challenge.

— PartiGen: Part-Aware Representations for Language-Conditioned Humanoid Motion Generation  (2610.13373 - Huang et al., 9 Oct 2026) in Appendix, Section 2, “Failure Cases and Limitations”