Being-H0.5 - A general-purpose robot model open-sourced by Lu Zongqing's team.

What is Being-H0.5?

Being-H0.5, developed by Lu Zongqing's team, is a general-purpose robot model that solves the problem of policy transfer between different robot hardware forms by leveraging prior human knowledge and unified action alignment. The model constructs the large-scale cross-morphological manipulation dataset UniHand-2.0 and designs a unified action space, enabling the model to learn general manipulation semantics and achieve shared policies among multiple robots. Combining MoT and MoF architectures, the model possesses powerful understanding and action generation capabilities, and ensures deployment stability in real-world environments through mechanisms such as MPG and UAC. Experiments show that Being-H0.5 performs excellently in long-range and dual-arm tasks, providing significant progress for the practical application of general-purpose robot models.

The main functions of Being-H0.5

• Cross-form controlThe model can share and transfer control strategies across different types of robots (such as robotic arms, dexterous hands, humanoid robots, etc.).
• Long-term task stabilityMaintaining stability and accuracy of actions in multi-step tasks to avoid error accumulation.
• Two-arm coordinationIt supports real-time collaboration in dual-arm tasks, ensuring the accuracy of action timing and spatial coupling.
• Generalization abilityIt maintains reliable performance across different environments, tasks, and hardware platforms, demonstrating strong adaptability.
• Real-world deployment stabilityThe problem of asynchrony between perception and control is solved by deployment mechanisms (such as MPG and UAC), ensuring stable operation on real robots.

The technical principle of Being-H0.5

• Unified State-Action SpaceThe state and actions of robots of different forms are mapped to a unified vector space, enabling the model to learn general manipulation semantics and action parameters that are not specific to hardware.
• Large-scale pre-trained dataset (UniHand-2.0)It integrates human hand manipulation data, robot control data, and visual language understanding data to provide the model with rich action priors and semantic alignment foundations.
• Hybrid architecture (MoT + MoF)The model adopts a Mixture-of-Transformers architecture, combining the Understanding Module (MoT) and the Action Generation Module (MoF) to improve the model's action generation and generalization capabilities.
• Action generation mechanism (Mixture of Flow)By using a shared layer and routing experts, both low-degree-of-freedom and high-degree-of-freedom actions can be handled simultaneously, avoiding negative migration.
• Deployment stability mechanism:
  • MPG (Manifold-Preserving Gating)Suppress unreasonable action output and maintain the stability of the action manifold.
  • UAC (Universal Async Chunking)This solves the problem of asynchronous frame rate perception and control frequency, ensuring the continuity of actions.
• Cross-morphological alignmentBy providing supervisory signals through human motion data, human interaction patterns can be directly applied to robot control, enhancing the model's generalization ability.

The project address for Being-H0.5

• GitHub repositoryhttps://github.com/BeingBeyond/Being-H
• HuggingFace model libraryhttps://huggingface.co/collections/BeingBeyond/being-h05
• arXiv technical paper: https://arxiv.org/pdf/2601.12993

Application scenarios of Being-H0.5

• Industrial AutomationIn factories, Being-H0.5 can control different types of robotic arms to complete complex assembly and handling tasks, improving production efficiency and flexibility.
• Home services: Used in home robots to complete multi-step tasks such as cleaning, tidying, and cooking, adapting to different home environments and items.
• Medical assistanceIn medical settings, it assists doctors in operating surgical instruments or in rehabilitation training, precisely controlling movements to ensure safety and stability.
• Logistics warehousingIt enables automated sorting, handling, and stacking of goods, adapting to items of different shapes and weights, and improving logistics efficiency.
• Educational ResearchAs a teaching tool, it helps students and researchers explore robot manipulation techniques and provides a cross-morphological experimental platform.