Gasgoo Munich-the China Academy of Information and Communications Technology (CAICT) has officially released the Industrial Embodied AI Robot Training Dataset 2.0. Developed in collaboration with Tsinghua University, Peking University, JD.com, ZTE, Estun, Hisense, Deta Intelligence, and Moja Robotics, the dataset is designed to serve as a high-quality foundation for training embodied AI models, verifying algorithms, and testing performance.
Image Credit: Deta Intelligence
Unlike traditional general-purpose datasets, Version 2.0 targets specific gaps in industrial embodied AI capabilities—autonomous perception, dynamic decision-making, and adaptive operation. It introduces three major breakthroughs:
First, it establishes the industry's first "trial-and-error" data system. By capturing the entire process—from deviation and problem identification to action adjustment and task completion—the system enables models to learn adaptive error correction within real industrial environments.
Second, it pioneers a "work-as-acquisition" data organization model. Partnering with Deta Intelligence, the initiative utilizes a globally unique full-body capture device centered on the human egocentric perspective. This approach synchronously collects, records, and converts worker actions, operational workflows, environmental perception, and human-machine interactions at actual workstations—significantly boosting the efficiency, authenticity, and scalability of data acquisition for embodied AI.
Third, it introduces a "real-scene data refining" development path. By embedding data collection teams directly into automotive factories and manufacturing frontlines, the project moves acquisition sites to actual production lines, workstations, and tasks. Frontline workers perform operations according to real production workflows, ensuring that data is sourced from the industrial field, trained within industrial scenarios, and applied directly to industrial use.
Deta Intelligence's head-mounted, full-panoramic data capture solution centers on the human egocentric perspective. It requires no virtual environment setup and does not interrupt production, allowing for the synchronous collection, recording, and conversion of worker operations, standardized workflows, workshop environmental perception, and human-machine interaction processes.
Deta Intelligence has completed the large-scale deployment of its proprietary full-panoramic capture devices. The system simultaneously collects full-body data on how hands, torso, legs, and feet interact with the environment, providing a solid data foundation for full-body coordination, the execution of long-sequence tasks, and cross-task generalization.
