Building a globally leading "Embodied Intelligent Super Supply Chain" JD.com (09618) released the industry's first full-body data full-link infrastructure.

One month after announcing the establishment of the world's largest embodied data collection center and supporting the trillion-dollar Siasun Robot & Automation industry ecosystem, JD.com (09618) has made another significant progress in the field of embodied intelligence. At the JD Embodied Intelligence Ecological Summit on April 16, JD.com announced the world's first full-chain embodied intelligent data infrastructure covering "acquisition, storage, labeling, training, evaluation, simulation, and testing." Self-developed ultra-high-definition acquisition terminal JoyEgoCam, embodied large model JoyAI-RA, and embodied intelligent data trading platform all made a stunning appearance. With the comprehensive use of the above infrastructure and its rich scene advantages in retail, logistics, industry, health, and other fields, JD.com is actively promoting the commercialization of embodied intelligence from laboratory research to large-scale business implementation, creating a "super smart embodied supply chain." 60,000 people, 10 million hours, the largest-scale embodied data collection operation in history Currently, embodied intelligence is accelerating in core areas such as industrial manufacturing, logistics warehousing, home services, medical rehabilitation, and more. However, the industry generally faces a core bottleneck - a severe lack of high-quality, real interactive data. Varying hardware standards, scattered data acquisition processes, disconnected labeling and training stages have led to the prominent issue of data silos. The collected data is chaotic, non-standard, making it difficult to be directly used for model training and unable to be compliant for circulation and trading. The industry consensus is clear: "The weakness of embodied intelligence hardware is fundamentally due to the lack of sufficient and good quality real data." Leveraging JD's accumulation in computing infrastructure, large model technology, and industrial scenarios, JD Cloud has launched the world's first full-chain embodied intelligent data infrastructure, bridging the entire process from data acquisition to model testing, refining raw chaotic data into high-value "data fuel" that drives model evolution. Based on this, JD will build the world's largest embodied intelligent data collection center, mobilizing up to 600,000 people for the "largest data collection operation in human history," accumulating 10 million hours of real human scene video data within two years, accelerating the iterative optimization of models in real scenes, and providing end-to-end data services for various industries such as Siasun Robot & Automation enterprises, embodied model research institutions, automakers, medical institutions, etc. Self-developed terminal JoyEgoCam allows real scene data to be "captured on the go" The first step in the embodied data chain is data collection. A major highlight of this summit is JD Cloud's self-developed wearable ultra-high-definition acquisition terminal JoyEgoCam. It guarantees the quality of source data from four dimensions: clarity, accuracy, portability, and stability. In terms of clarity, JoyEgoCam is equipped with a 4K HD camera that supports 60 fps frame rate and 130-degree ultra-wide-angle shooting, enabling millisecond-level capture of motion details, accurately recording subtle operations in various scenes. In terms of accuracy, the reprojection error is less than 0.2 pixels, combined with JD Cloud's self-developed stereo correction technology, it can realistically reproduce the spatial stereoscopic sense of the operating scene. In terms of portability, the entire device weighs only 220 grams, lighter than a regular smartphone, comfortable to wear. In terms of stability, it has an internally built-in automotive-grade 6-axis IMU and a fusion unit of multiple sensors, ensuring stable tracking and shooting even in extreme shaking scenes. JoyEgoCam can be worn and used for data collection in various scenes such as logistics, retail, medical, and home, allowing ordinary people to complete professional-level data collection, solving the pain points of "data not genuine, not accurate" from the source. Full-chain processing platform, increasing model training efficiency by 3.5 times After data collection, it enters the upload and processing stage. JD Cloud achieves one-click cloud storage of videos through task, personnel, and equipment visual management and SaaS deployment, significantly improving collection efficiency and reducing costs. After data is converged into the AI data lake platform, it automatically completes cleaning, alignment, transformation, and pre-labeling with PB-level throughput capacity, transforming it into a standard training set. The JoyBuilder simulation platform generates high-fidelity simulated data in batches, efficiently transforming and generalizing human-operated data simulated operation data real machine operation data in one stop, filling the long-tail scenarios, and doubling the evaluation efficiency. Governed data is aggregated on the JoyBuilder model development platform, achieving data "unboxing and training," model "one-click deployment," increasing model training efficiency by 3.5 times, significantly lowering the threshold for developing VLA large models. The self-developed AI operator matrix runs through the process, covering key steps such as distortion correction, semantic description, deep reconstruction, refining high-value training materials. Currently, JD daily processes tens of thousands of data entries, with a data processing efficiency of up to 95%, and overall processing costs reduced by 60%. Embodied large model JoyAI-RA, globally leading in real machine success rate With the full-chain infrastructure in place, JD has built an ecological loop of "data acquisition - model training - data optimization." Based on the core-trained self-acquired data, JD's embodied large model JoyAI-RA achieves a success rate of 73.5% in real machine experiments, surpassing SOTA models such as pi0.5. The model feeds back to data labeling and collection, forming a "stronger with use, more cost-effective" flywheel. The higher the data quality, the higher the efficiency and accuracy of model iteration; the more mature the model, the lower the cost and higher the quality of data acquisition and labeling, achieving mutual empowerment and collaborative evolution of data and models, further enhancing JD's technical accumulation and industrial insight in the field of embodied intelligence. Data trading platform and open community, making data "flowable and usable" A noteworthy release is the simultaneous launch of the JD embodied intelligent data trading platform, bringing together diverse multimodal data resources from JD's rich business scenarios, supporting collaboration among data providers, developers, and application parties, and ensuring compliant circulation through end-to-end security audits. The platform's first batch includes 2,000 hours of high-precision labeled data set to open up compliant trading channels for embodied intelligent data. JD calls for more industry partners to join and build an "embodied intelligent data circle of friends." The upcoming technical community will also gather high-quality industry resources, promoting exchange, innovation, and industrial collaboration. JD concentrated on releasing embodied intelligence-related products and infrastructure, not only filling the industry gap in data acquisition, processing, and circulation but also building a complete industry chain of "hardware acquisition - data processing - model training - simulation testing - compliant trading - ecological co-construction." Leveraging the foundation of data and models, JD's layout in the field of embodied intelligence is rapidly expanding: JD JoyInside embodied intelligence has deep cooperation with nearly 200 brands in home appliances, home furnishings, Siasun Robot & Automation, toys, injecting "wisdom and soul" into smart terminals; JD Retail will help Siasun Robot & Automation brand partners achieve cumulative sales exceeding 10 billion by the year 2026; JD Logistics continues to build a Siasun Robot & Automation after-sales service ecosystem, with Siasun Robot & Automation ambulances serving the Chinese and overseas markets, and the scale of professional engineers expanding to over ten thousand; JD Industrials is creating a one-stop industrial supply chain technology and service to achieve 100% coverage of Siasun Robot & Automation manufacturing materials... In the future, JD will continue to transform its accumulation of supply chain, technology, and scenario into industrial service capabilities, promoting the deep landing of embodied intelligence in various industries, driving the development of a trillion-level embodied intelligence ecosystem.