京东将建全球最大具身数采中心 未来五年买300万台机器人

Key Information

1. JD plans to build the world's largest embodied AI data collection center, aiming to collect more than 10 million hours of human real-world video data and more than 1 million hours of robot hardware operating data within two years. It already covers more than 100 sub-scenarios and over 100,000 real work tasks.

2. JD is building a super AI supply chain through six modules: cloud, data, models, devices, scenarios, and chain. It emphasizes a self-developed foundation plus an open ecosystem, validating internally first before opening to the industry.

Practical Takeaways

1. The data closed loop advances through data collection, model training, and effect validation, with JoyAI-Sim used for simulation evaluation. Simulation-to-real consistency is about 90%, and efficiency is 3.2 times that of real machines.

2. JD Cloud offers the JoyScale compute platform, the JoyBuilder training platform, a self-developed inference engine, and JoyMaaS model services. Inference costs for typical tasks are reduced by up to 80%.

3. In the next five years, JD Logistics plans to procure 3 million robots, 1 million unmanned vehicles, and 100,000 drones, and will open up scenarios, procurement demand, and sales channels.

Brands

1. Product R&D and scenario validation: JD has more than 3,800 warehouses, more than 200,000 partner offline retail pharmacies, and more than 10,000 stores including JD MALL, 7Fresh, and JD Auto Care. These provide real environments and continuous feedback for robots to learn product handling, warehouse operations, and home services.

2. Channels and market opportunities: JD will open up devices, real-world scenarios, procurement demand, and sales channels to move robots from products to market. In the next five years, JD Logistics plans to procure 3 million robots, 1 million unmanned vehicles, and 100,000 drones.

3. Consumer trend observation: the article argues that AI is moving from the digital world to the physical world, from understanding and generating information to sensing the environment, making judgments, and executing tasks. It emphasizes making life better and bringing AI into every home, indicating that embodied AI is extending into home services and life experiences.

4. Brand partnership implications: JD is building a physical AI collaborative network as a connector and operator. Brands can watch its scenario opening, data trading platform, and supply chain resources, but the article does not provide specific pricing or price competition information.

Sellers

1. Growth market and opportunity signals: JD plans to collect more than 10 million hours of human video and more than 1 million hours of robot data within two years, and already covers more than 100 sub-scenarios in production and services. In the next five years, it plans to procure 3 million robots, 1 million unmanned vehicles, and 100,000 drones, and will open up procurement demand and sales channels.

2. Cooperation approach: JD is building an embodied AI data trading platform to promote data circulation and reuse under authorization and security boundaries. It has integrated high-quality datasets at the tens-of-thousands scale across multiple scenarios and works with dozens of data companies. It has launched the JD Physical AI Acceleration Plan, with the first phase focused on embodied intelligent robots and investment in six directions: data, bases, manufacturing, sales, logistics, and services.

3. Risk and compliance notes: data collection is carried out strictly in accordance with laws and regulations to ensure legality, compliance, safety, and trustworthiness. Partners should pay attention to authorization and security boundaries during cooperation.

4. Key learnings: start with business problems, then refine algorithms, accumulate data, and iterate models. Validate within JD's own business first, then open to the industry.

Factories

1. Production and design demand: in the next five years, JD plans to procure 3 million robots, 1 million unmanned vehicles, and 100,000 drones, and has launched the Physical AI Acceleration Plan. The first phase invests in six directions: data, bases, manufacturing, sales, logistics, and services, unlocking demand on the manufacturing side.

2. Business opportunities: JD will open up devices, real-world scenarios, procurement demand, and sales channels. Factories can seek cooperation around embodied robots, unmanned vehicles, drones, and supply chain supporting capabilities. Zhiyuan's embodied robots have entered JD Logistics' Zhilang Warehouse, testing product boundaries and improving reliability in real operations.

3. Digitalization implications: JD uses cloud, data, models, devices, scenarios, and chain to build a super AI supply chain and trains models with real-scenario data. Its data service system covers collection, storage, annotation, training, evaluation, simulation, and testing. Simulation evaluation is about 90% consistent with real machines, and efficiency is 3.2 times that of real machines.

4. Implementation path: robots need to move from single-point operations to process collaboration and integrate into industry processes to achieve scale.

Service Providers

1. Industry trend: competition in physical AI is not only single-point competition in models and robot hardware, but competition in system capabilities. AI is moving from understanding and generating information to sensing the environment, making judgments, and executing tasks, and a continuous supply of real operating data has become critical.

2. Customer pain points: enterprises need fast training, stable operation, and iterability. Embodied AI customers face insufficient real-world scenario data, long training cycles, and high data preprocessing investment.

3. Solutions: JD Cloud provides the JoyScale compute platform for compute pooling, heterogeneous scheduling, and software-hardware co-optimization. It has jointly built a fully domestic 10,000-GPU cluster and is planning a 100,000-GPU cluster. JoyBuilder provides end-to-end model training and connects training, evaluation, and simulation into a closed loop. A self-developed inference engine reduces inference costs for typical tasks by up to 80%. JoyMaaS aggregates more than 100 mainstream open-source models and self-developed models.

4. Service case: after an embodied AI foundation-model company accessed JD Cloud data, its success rate in operating unknown objects improved significantly, while its training cycle and data preprocessing investment dropped sharply.

Marketplace Platforms

1. New platform practices: JD is building a super AI supply chain around six modules: cloud, data, models, devices, scenarios, and chain. It plans to build the world's largest embodied AI data collection center and an embodied AI data trading platform, integrating high-quality datasets at the tens-of-thousands scale across multiple scenarios.

2. Platform investment and cooperation: JD opens scenarios and supply chain resources to industry partners, including devices, real-world scenarios, procurement demand, and sales channels. It has cooperated with dozens of data companies, and JD Cloud serves UBTECH and other embodied AI industry customers.

3. Operations management: JD emphasizes a self-developed foundation plus an open ecosystem, validating within its own business first before opening to the industry. Data collection follows laws and regulations to ensure legality, compliance, safety, and trustworthiness, and data trading clearly defines authorization and security boundaries.

4. Risk avoidance: physical AI is competition in system capabilities, requiring integration of compute, training, inference services, and application scenarios. Players should avoid relying only on single points such as models or hardware, and should advance the world's largest physical AI collaborative network as a connector and operator.

Researchers

1. New industry developments: JD proposes a super AI supply chain composed of cloud, data, models, devices, scenarios, and chain. It plans to build the world's largest embodied AI data collection center, collecting more than 10 million hours of human real-world video and more than 1 million hours of robot hardware data within two years, and already covers more than 100 sub-scenarios and over 100,000 real work tasks.

2. New questions: competition in physical AI is competition in system capabilities, not only single-point competition in models and robot hardware. A continuous supply of real operating data is key for AI to move from information processing to sensing, judgment, and execution. Data collection and circulation must be legal, compliant, safe, and trustworthy, with clear authorization boundaries.

3. Business model: JD adopts a self-developed foundation plus an open ecosystem, validating internally first and then opening to the industry. Through an embodied AI data trading platform, it promotes data circulation and reuse, and builds a physical AI collaborative network and a physical-world operations center as a connector and operator.

4. Policy and regulatory implications: the article emphasizes strictly carrying out data collection in accordance with laws and regulations to ensure legality, compliance, safety, and trustworthiness. The data trading platform operates under authorization and security boundaries, and can serve as an observation sample for data governance and industry collaboration.