Being-H0.5: Scaling Human-Centric Robot Learning for Cross-Embodiment Generalization

TL;DR

Being-H0.5 introduces a human-centric paradigm for cross-embodiment robot learning by unifying human demonstrations and robotic data into a common state-action space and a single sequence modeling objective. Leveraging UniHand-2.0, the largest embodied pre-training corpus to date with over 35,000 hours and 120B tokens across 30 embodiments, the model employs a Mixture-of-Flow architecture, Manifold-Preserving Gating, and Universal Async Chunking to achieve robust real-time control across diverse hardware. Empirical results demonstrate SoTA performance on LIBERO (≈98.9% average) and RoboCasa (≈53.9%), plus strong cross-embodiment transfer across five real robots and multiple simulations, including embodiment-level zero-shot transfer. The work provides a scalable path toward generalist robotics by embedding transferable motor priors from human data into a unified action language, enabling deployment across heterogeneous morphologies with limited target-domain data.

Abstract

We introduce Being-H0.5, a foundational Vision-Language-Action (VLA) model designed for robust cross-embodiment generalization across diverse robotic platforms. While existing VLAs often struggle with morphological heterogeneity and data scarcity, we propose a human-centric learning paradigm that treats human interaction traces as a universal "mother tongue" for physical interaction. To support this, we present UniHand-2.0, the largest embodied pre-training recipe to date, comprising over 35,000 hours of multimodal data across 30 distinct robotic embodiments. Our approach introduces a Unified Action Space that maps heterogeneous robot controls into semantically aligned slots, enabling low-resource robots to bootstrap skills from human data and high-resource platforms. Built upon this human-centric foundation, we design a unified sequential modeling and multi-task pre-training paradigm to bridge human demonstrations and robotic execution. Architecturally, Being-H0.5 utilizes a Mixture-of-Transformers design featuring a novel Mixture-of-Flow (MoF) framework to decouple shared motor primitives from specialized embodiment-specific experts. Finally, to make cross-embodiment policies stable in the real world, we introduce Manifold-Preserving Gating for robustness under sensory shift and Universal Async Chunking to universalize chunked control across embodiments with different latency and control profiles. We empirically demonstrate that Being-H0.5 achieves state-of-the-art results on simulated benchmarks, such as LIBERO (98.9%) and RoboCasa (53.9%), while also exhibiting strong cross-embodiment capabilities on five robotic platforms.

Figures (14)

• Figure 1: Being-H0.5 at a Glance. We scale human-centric robot learning with Being-H0.5 toward cross-embodiment generalization. We introduce UniHand-2.0, a large-scale corpus exceeding 35,000 hours that spans both cross-Embodiment physical control and general visual-text understanding. Building on this data, we unify human hand motion and diverse robot embodiments with a Unified Action Space, and train all heterogeneous supervision through unified sequence modeling under a single framework. This yields a single foundation model that can perceive, describe, and act within one framework, enabling robust cross-embodiment generalization and real-world deployment across diverse robots and tasks. We empirically deploy a single checkpoint of Being-H0.5 to control PND Adam-U, Franka+Inspire, Unitree G1, BeingBeyond D1, and LeRobot SO-101 to accomplish diverse tasks.
• Figure 2: Overview of UniHand 2.0. UniHand 2.0 is our large-scale pre-training recipe for human-centric robot learning, comprising 35K hours of multimodal data from three complementary sources. 1) Human Demonstration with diverse scenes, multitask for motion alignment, and multi-granularity semantics. 2) Robot Manipulation spanning 30+ embodiments with multiple observation views and heterogeneous control signals. 3) Vision–Text Understanding covering general VQA, 2D spatial grounding & affordance, and task planning & reasoning.
• Figure 3: Comparison of training scale and embodiment diversity. Stacked bars represent hours of training data (left axis); hatched bars represent embodiment counts (right axis). UniHand-2.0 represents the largest and most diverse VLA pre-training recipe to date, totaling 35,000 hours of multimodal data. This includes 16,000 hours of human data, 14,000 hours of robot data across 30 embodiments, and 5,000 equivalent hours of VLM data.
• Figure 4: Statistics of UniHand-2.0. (Left) Ratio of simulation vs. real-world data. We maintain a balanced ratio with simulation data at 26%, while the widely used Open X-Embodiment (OXE) and AgiBot World datasets account for 3.1% and 3.0%, respectively. (Right) Training sources and scale. Human, robot, and visual-language ("VLM") data consist of 16K hours (25.6B tokens), 14K hours (45.7B tokens), and 5K equivalent hours (50.2B tokens), respectively. These sources are curated to maintain a comparable scale for balanced pretraining.
• Figure 5: An overview of our data collection system UniCraftor.