How JD.com Is Building an Open, Full-Stack Physical AI Ecosystem

JD.com is positioning AI as industrial infrastructure: open enough for partners and developers to use, but grounded in the company’s own retail, warehousing, logistics, health, and manufacturing operations. Founder and chairman Richard Liu has described technological barriers as a form of exploitation and said JD does not want its AI capabilities to serve only one company or country. The practical expression of that view is a full-stack strategy that links models, data, hardware, deployment, and service. 12

The distinction matters. JD is not simply releasing a chatbot or a collection of model weights. It is trying to build a repeatable path from real-world data collection to model training, robot manufacturing, operational testing, and commercial use. That could lower the cost and complexity of adopting physical AI, although the long-term economics and openness of the ecosystem have not yet been independently established.

The blueprint starts with heavy investment and selective openness

JD reported that its first-half 2026 R&D spending increased 53.2% year over year, with investment directed toward the JoyAI model family, embodied intelligence, physical-AI infrastructure, and industrial applications. The company has also said its full-stack, self-developed AI will be opened to global partners. 23

That openness is taking several forms. JD has released selected models through Hugging Face and GitHub, made digital-human livestreaming available free to more than 80,000 merchants, and opened the EgoLive human-view dataset to industry users. 238

This is best understood as an ecosystem strategy rather than a claim that every component is fully open on identical terms. Model weights, training materials, datasets, deployment systems, and commercial services can each have different access conditions. For users, the important question is not only whether a model is labeled “open source,” but whether the surrounding tools and data make it practical to build and operate an application.

Real operations provide the data advantage

Physical AI needs more than internet-scale text. Robots must interpret objects, spaces, motion, timing, safety constraints, and the consequences of their actions. JD’s advantage, in its own account, is more than two decades of experience across warehousing, retail, and logistics, giving it access to operational environments in which these problems occur repeatedly. 234

The company plans to collect more than 10 million hours of first- and third-person real-world video over two years, with participation spanning retail, warehousing, logistics, industrial work, household settings, and other scenarios. It is also building what it describes as the world’s largest embodied-data collection center and has established an embodied-data community in Suqian. 1562930

EgoLive is the initial public layer of this effort. JD has described it as the industry’s largest human-view dataset; reports on its first release describe roughly 2,000 hours covering more than 300 operational tasks. High-precision datasets have also been made available on a targeted basis to universities, developers, and ecosystem companies. These scale descriptions come from JD or industry reporting and should be treated as claims rather than independently verified benchmarks. 45630

The intended data flywheel is straightforward:

1. Real workers and machines generate video and operational data.
2. That data helps train models for perception, planning, and interaction.
3. Improved models are tested in warehouses, delivery networks, factories, and service settings.
4. Deployments generate new examples, exceptions, and performance feedback.
5. The resulting capabilities can be packaged for more users and industries.

This is why JD’s strategy depends on operating physical businesses. An open model can reduce the barrier to experimentation, but reliable physical AI also requires access to real tasks and the infrastructure needed to measure whether a system works.

JoyAI is a portfolio of building blocks, not one monolithic model

JD’s open AI push spans several types of models. JoyAI-Echo focuses on long-form audio-video generation, JoyAI-VL-Interaction on real-time vision-language interaction, and JoyAI-Video-Edit on streaming video editing. 571118

JoyAI-VL-Interaction is particularly relevant to physical environments because it is designed to process ongoing video rather than only answer questions about a static image. Its release includes the model, training recipe, time-aligned interaction data, and a deployable real-time streaming stack. The project’s technical documentation describes an 8-billion-parameter vision-first system that can process continuous video with sub-second latency. 171923

That packaging lowers the engineering burden for developers who want to build around continuous camera input. It also reflects JD’s broader approach: release reusable components while retaining an advantage in the real-world environments where those components are trained, evaluated, and integrated.

RoboBase extends the strategy from software to manufacturing

Model access alone does not create a physical-AI industry. Robots need components, assembly, testing, integration, maintenance, and local support. JD’s RoboBase initiative is intended to supply that missing industrial layer.

The first RoboBase facility began construction in Guangzhou in 2026. JD says the project will support a research-to-manufacturing-to-application-to-service cycle and that it plans to establish more than 80 robot bases across China over five years. 36810

That structure could make the company more than an AI application provider. It would position JD as an infrastructure and services partner for robotics companies, combining supply-chain access, model capabilities, data collection, testing environments, production capacity, and after-sales support. The proposal is ambitious, however, and the available reporting describes plans and company positioning more clearly than independently audited operating results.

Logistics deployments are the proof-of-repeatability test

JD is using its logistics network as the main demonstration environment for physical AI. Company and industry reports cite the following examples:

• The Zhilang goods-to-person system has been deployed across more than 60 warehouses, including overseas markets.
• The Yilang dual-arm robot can reportedly recognize, grasp, and stack an item in about 10 seconds.
• More than 1,000 autonomous delivery vehicles operate across more than 20 Chinese provinces.
• A drone delivery network in Sichuan covers 78 villages, with some deliveries completed in about seven minutes. 81124

These examples cover warehousing, picking, handling, sorting, and last-mile delivery. Their strategic importance is not merely that they demonstrate automation. They are intended to show that a model-and-robot system can be integrated into repeatable workflows across multiple locations rather than remaining a laboratory prototype. 24

The evidence is still primarily company-reported or based on industry coverage, so the strongest conclusion is limited: JD has established a substantial set of claimed deployments and is using them to support its industrialization narrative. Whether the systems deliver consistent cost, safety, uptime, and labor outcomes at scale remains a question for further verification.

The ecosystem reaches beyond JD’s warehouses

JD says its industrial AI agents serve more than 3,000 manufacturers and can reduce data-governance work from months to hours. It also says its JoyAvatar digital-human tools have been used by more than 80,000 merchants, while second-quarter live-streaming accounts tripled year over year. The Aidol Creation Camp has helped bring more than 50 AI-hardware products to market, according to company and industry reports. 28

These initiatives broaden the target customer base. Merchants can use digital presenters, manufacturers can apply industrial agents, universities and developers can work with datasets, and hardware startups can gain access to supply-chain and application support. In theory, each new participant adds more deployment feedback and expands the market for the underlying tools.

That is the difference between an internal automation program and an industrial platform strategy. JD is attempting to distribute AI capabilities outward while capturing value from the surrounding services: improved fulfillment, lower operating costs, hardware production, integration, merchant growth, and industrial productivity.

What JD’s model still needs to prove

JD’s blueprint is coherent: invest heavily in the stack, open selected capabilities, collect data from real work, train models for physical tasks, manufacture and test robots, and deploy them through an existing operating network. It also fits a broader shift in physical AI, where the competitive unit is increasingly a complete system rather than a model in isolation. 3431

But openness does not automatically mean low cost, and deployment does not automatically mean durable productivity gains. Independent evaluation will be needed on several points:

• How broadly partners can use the models, data, and deployment software
• Whether the datasets have sufficient documentation, permissions, and governance safeguards
• How well the systems transfer across warehouses, countries, industries, and hardware platforms
• Whether robot deployments reduce total operating costs after integration, maintenance, and supervision
• How third-party developers and manufacturers share in the value created by the ecosystem

For now, JD’s clearest strategic bet is that AI becomes more useful—and potentially more accessible—when it is connected to real physical work. Its open models and datasets are the visible layer. The deeper proposed moat is the full loop underneath: operational data, tested workflows, manufacturing capacity, deployment scale, and service infrastructure.