At the 27th China Hi-Tech Fair held in Shenzhen from November 14-16, 2025, Wu Bangyi, Chief Data Officer of Tianyu Digital Technology (Dalian) Group Co., Ltd., shared insights on the development of embodied AI during the China Hi-Tech Forum.
Wu highlighted that embodied AI is evolving from "passive perception" to "active perception," driving rapid industry expansion. By the end of 2024, over 451,700 Chinese companies were engaged in this field, with numbers continuing to rise. He projected that by 2029, China would capture half of the global market share in embodied AI and robotics.
However, Wu pointed out two major industry bottlenecks: 1. **Lack of platforms and standards** – Humanoid robots vary in design, leading to fragmented training methods and weak model generalization. 2. **Scarcity of 3D data** – While 2D text models are trained on 50-80TB of data, 3D spatial data remains extremely limited, hindering the development of spatial intelligence models.
To address these challenges, Tianyu Digital Technology focuses on infrastructure through its "ABC" framework: - **A (Assets)** – The BehavisionPro platform has accumulated over 1.5 million 3D datasets and 650,000 multimodal datasets, forming a hierarchical data ecosystem. - **B (Behavior)** – Enhances decision-making models for robotic applications. - **C (Client)** – Aims to create an "Android-like" platform for embodied AI, supporting diverse enterprises.
Wu also discussed the high costs of real-world robot training. For example, teaching a robot to retrieve an apple from a fridge requires 50-100 sets of motion data, translating to over 100,000 data points per action. To mitigate costs, Tianyu leverages simulation data and repurposes motion-capture datasets from its metaverse projects.
Additionally, the company has developed specialized 3D articulated datasets, enabling robots to understand object interactions (e.g., hinges on a trash can or microwave). Its Behavior platform integrates Large Language Models (LLMs) and Vision Language Models (VLMs) for long- and short-term task planning, improving robotic adaptability in real-world scenarios.
Wu concluded by envisioning a future of diverse robotic forms across industries, powered by modular AI systems.
Comments