UniT: Toward a Unified Physical Language for Human-to-Humanoid Policy Learning and World Modeling

arXiv cs.RO / 4/22/2026

📰 NewsSignals & Early TrendsIdeas & Deep AnalysisModels & Research

共有:

Key Points

UniT proposes a unified “physical language” to transfer learning from humans to humanoid robots despite large kinematic differences between embodiments.
It uses a tri-branch cross-reconstruction approach (action→vision anchoring, vision→action filtering, plus a fusion module) to create an embodiment-agnostic discrete latent space of physical intents.
The framework is validated in two settings: VLA-UniT for humanoid policy learning using diverse human data, and WM-UniT for world modeling and human-to-humanoid action transfer.
Results report improved data efficiency, robust out-of-distribution generalization, and notable zero-shot task transfer, alongside enhanced action controllability for humanoid video generation.
The authors claim empirically aligned cross-embodiment representations, supported by t-SNE visualizations showing convergence of human and humanoid features into a shared manifold.

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 synergies these purified modalities into a shared discrete latent space of embodiment-agnostic physical intents. We validate UniT across two paradigms: 1) Policy Learning (VLA-UniT): By predicting these unified tokens, it effectively leverages diverse human data to achieve state-of-the-art data efficiency and robust out-of-distribution (OOD) generalization on both humanoid simulation benchmark and real-world deployments, notably demonstrating zero-shot task transfer. 2) World Modeling (WM-UniT): By aligning cross-embodiment dynamics via unified tokens as conditions, it realizes direct human-to-humanoid action transfer. This alignment ensures that human data seamlessly translates into enhanced action controllability for humanoid video generation. Ultimately, by inducing a highly aligned cross-embodiment representation (empirically verified by t-SNE visualizations revealing the convergence of human and humanoid features into a shared manifold), UniT offers a scalable path to distill vast human knowledge into general-purpose humanoid capabilities.