The scaling of humanoid foundation models faces a critical bottleneck: the scarcity of robotic data. While vast amounts of human-centric egocentric data exist, the kinematic differences between humans and humanoids present a significant hurdle for effective transfer. This challenge is addressed by UniT (Unified Latent Action Tokenizer via Visual Anchoring), a novel framework designed to forge a universal physical language for human-to-humanoid transfer.
Forging a Universal Physical Language
UniT operates on the principle that disparate kinematic structures share universal visual consequences. The framework employs a tri-branch cross-reconstruction mechanism. Actions are used to predict vision, effectively anchoring kinematics to physical outcomes. Concurrently, vision reconstructs actions, filtering out irrelevant visual confounders. A crucial fusion branch then synergizes these purified modalities into a shared, discrete latent space representing embodiment-agnostic physical intents. This approach creates a common ground for understanding and transferring actions across different embodiments.
Dual-Paradigm Validation: Policy Learning & World Modeling
The efficacy of UniT is demonstrated across two key paradigms. In Policy Learning (VLA-UniT), the prediction of these unified tokens allows for the effective leveraging of diverse human data. This results in state-of-the-art data efficiency and robust out-of-distribution (OOD) generalization on both simulated humanoid benchmarks and real-world deployments, notably achieving zero-shot task transfer. For World Modeling (WM-UniT), UniT aligns cross-embodiment dynamics by conditioning on unified tokens. This enables direct human-to-humanoid action transfer, ensuring that human data seamlessly translates into enhanced action controllability for humanoid video generation. Empirical verification via t-SNE visualizations confirms a highly aligned cross-embodiment representation, with human and humanoid features converging into a shared manifold.