Recently, Hefei Embodied Data Operation and Management Co., Ltd. completed a special property right registration at the Anhui Data Exchange. The self-collected and self-generated 2025–2026 Hefei Intelligent Robot Universal Object Grasping Dataset of the company has officially obtained its registration certificate. This marks Anhui’s first registered property right for a multimodal grasping dataset tailored to embodied intelligent robots.

What is this dataset for? Simply put, it serves as teaching materials for robots to “learn to perform tasks”. “Robots need to recognize and grasp various items such as pens, cups and storage boxes through repeated practice. They capture visual footage via cameras and execute movements with grippers, recording all information generated throughout the process—including visual frames, command texts and wrist positioning—to form a standardized dataset,” explained Du Wenhe, Technical Director of Hefei Embodied Data Operation and Management Co., Ltd.

The registered dataset was collected in real-world scenarios using LindenBot general-purpose robots. It covers full workflow scenarios including grabbing, placing, sorting and storage, with comprehensive and multi-dimensional data coverage: robotic task text instructions, synchronized video streams from three cameras (head-mounted and left/right wrist-mounted), and real-time pose data of the end grippers.

Du Wenhe noted that earlier robot systems could only process static images or verbal commands in isolation. “Multimodality means this dataset supports multiple information formats including text instructions, visual imagery and verbal descriptions. Large models trained on such data become far more intelligent—able to perceive visuals, comprehend voice commands and carry out physical operations autonomously. Processed through standardized collection, algorithmic cleaning and calibration, as well as manual verification, the dataset boasts high industrial application value and supports three core technological tracks:

1. Training vision-language-action multimodal large models for wheeled and robotic-arm robots, enabling robots to interpret instructions, identify objects and complete operations independently to achieve genuine autonomous control;

2. Pre-training world models to build foundational vision-language models that serve as the training base for advanced vision-language-action models;

3. Facilitating data fusion across heterogeneous robots for unified large-model training.

By clarifying the legal ownership of embodied robotic data, this registration removes practical barriers hindering efficient circulation and utilization of vertical industry data. The property right certificate clearly defines data ownership and authorized users, and enables legal data trading. It provides a legitimate, reliable data foundation for the research and development of embodied intelligence and general robotic large models across Anhui Province.

At present, Hefei Embodied Data Operation and Management Co., Ltd. operates a standardized robot data collection training base spanning over 5,000 square meters, generating 2 million compliant data entries annually. Moving forward, the company will continue to advance standardized data supply with its multi-scenario training grounds, steadily drive the transformation of data resources into data assets and data capital, and build a compliant, high-quality, reusable public data base for embodied intelligence. It will empower and boost the R&D of general robotic large models, as well as the industrial transformation, upgrading and innovative growth of the robotics sector in Hefei and the whole of Anhui Province.

Source: Hefei Daily