论文概要
Research area: Machine Learning Authors: Boyu Chen, Yi Chen, Lu Qiu, Jerry Bai, Yuying Ge, Yixiao Ge arXiv: 2604.19734
Abstract
Scaling humanoid foundation models is bottlenecked by the scarcity of robotic data. While massive egocentric human data offers a scalable alternative, bridging the cross-embodiment chasm remains a fundamental challenge due to kinematic mismatches. We introduce UniT (Unified Latent Action Tokenizer via Visual Anchoring), a framework that establishes a unified physical language for human-to-humanoid transfer. Grounded in the philosophy that heterogeneous kinematics share universal visual consequences, UniT employs a tri-branch cross-reconstruction mechanism: actions predict vision to anchor kinematics to physical outcomes, while vision reconstructs actions to filter out irrelevant visual confounders. Concurrently, a fusion branch synergizes these purified modalities into a shared discrete latent space of embodiment-agnostic physical intent.
The framework is validated under two paradigms:
1. Policy learning (VLA-UniT): By predicting these unified tokens, the model effectively exploits diverse human data, achieving state-of-the-art data efficiency and strong out-of-distribution (OOD) generalization on humanoid simulation benchmarks and real-world deployment, notably demonstrating zero-shot task transfer. 2. World modeling (WM-UniT): Using the unified tokens as conditioning, dynamics are aligned across embodiments, enabling direct human-to-humanoid action transfer. This alignment ensures human data can be seamlessly converted to enhance action controllability in humanoid video generation.
Finally, by inducing highly aligned cross-embodiment representations — empirically revealed by t-SNE visualization showing human and humanoid features converging onto a shared manifold — UniT offers a scalable path for distilling massive human knowledge into general-purpose humanoid capabilities.
---
*Auto-collected on 2026-04-23. Originally posted on zhichai.net.*