京东披露具身智能数采进展:覆盖百余细分场景、采集超 10 万真实作业任务JD.com reveals embodied AI data collection: 100k+ real tasks across 100+ scenarios


AI free token/AI免费积分

IT之家 11 月 4 日消息,今日在 2026 年世界互联网大会乌镇峰会机器人与具身智能论坛上,京东集团工艺委员会主席曹鹏发表演讲,分享了京东推动 AI 从数字世界走向物理世界的最新实践。当下,京东具身数采已覆盖 100 余个细分生产服务情形,采集超 10 万真实作业任务。

图片

曹鹏表示,物理 AI 的竞争,不只是模型和机器人本体的单点竞争,更是体系能力的竞争。京东将坚持“自研底座 + 开放生态”,把经过真实业务验证的能力向产业开放,让 AI 转化为提高产业效率、改善生活感受的生产力。

IT之家注意到,当下,京东 AI 已应用于零售、物流、健康、工业等超过 3000 个情形。另外,京东具备 3800 多个仓库、超过 20 万家合作线下零售药房,以及超过 1 万家京东 MALL、七鲜、养车等各类门店。多元的业务情形,为机器人学习商品操作、仓储作业和家庭服务等任务给出了真实环境与持续反馈。

图片

为将情形资源转化为可训练的数据,京东正在推进具身智能数据采集中心建设,计划两年内累计采集超过 1000 万小时高质量人类真实情形视频数据,以及超过 100 万小时机器人本体运行数据,相关采集将严格依法依规开展,确保合法、合规、安全、可信。当前,已覆盖 100 余个细分生产服务情形、采集超 10 万真实作业任务。

围绕数据利用,京东云自研 JoyEgoCam 等采集终端,并构建覆盖“采、存、标、训、评、仿、测”的具身数据服务体系。另外,京东还推出 JoyAI-Sim 物理仿真引擎,连接人类操作数据、仿真环境和机器人实测,兼容数据扩增与策略验证。在已完成的测试中,仿真实测与真实实测结果的一致性约为 90%,实测效率达到真机实测的 3.2 倍。京东自研具身操作基础模型 JoyAI-RA 的测试表明,随着人类实操视频数据规模扩大,模型真机任务得分稳步增长。

在算力维度,京东云 JoyScale AI 算力平台借助算力池化、异构调度和软硬协同优化,达成多元算力统一管理与高质调度。当下,京东云已联合国内合作伙伴构建纯国产万卡集群,并规划建设十万卡集群,为大规模模型训练和物理 AI 应用给出支撑。

在训练维度,JoyBuilder 模型开发平台为具身模型训练给出端到端服务,优化多模态数据读取与训练效率,打通训练、实测和仿真闭环,让企业将更多精力投入模型与业务创新。

在推理与模型服务维度,京东云自研推理引擎,大幅提高高并发情形下的吞吐与资源利用效率,典型任务推理成本最高降低 80%。一站式大模型服务平台 JoyMaaS 已汇聚 100 余个主流开源模型与京东自研模型,给出统一接入、企业级管理和成本治理能力,降低企业开发与应用门槛。

在应用端,京东正推动 AI 与机器人融入供应链全流程。例如,将来五年,京东物流计划采购 300 万台机器人、100 万台无人车和 10 万架无人机。智元的具身机器人已进入京东物流智狼仓,双方在真实作业中检验机型边界、提高可靠性,推动机器人从单点操作走向流程协同。

图片

本文来源:https://www.ithome.com/1/022/356.htm
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

IT Home, November 4 news: Today at the Robotics and Embodied Intelligence Forum of the 2026 World Internet Conference Wuzhen Summit, Cao Peng, Chairman of JD's Technology Committee, delivered a speech sharing JD's latest practices in driving AI from the digital world into the physical world. Currently, JD's embodied data collection has covered over 100 specific production and service scenarios, with over 100,000 real-world tasks collected.

图片

Cao Peng stated that the competition in physical AI is not just about single-point competition in models and robot hardware, but more importantly, competition in system-level capabilities. JD will adhere to a "self-developed foundation + open ecosystem" strategy, opening up capabilities validated through real business operations to the industry, turning AI into a productive force that improves industrial efficiency and enhances life quality.

IT Home noted that JD's AI is currently applied in over 3,000 scenarios across retail, logistics, healthcare, and industry. In addition, JD operates over 3,800 warehouses, more than 200,000 partner offline retail pharmacies, and over 10,000 stores including JD MALL, 7Fresh, and JD Auto. These diverse business scenarios provide a real environment and continuous feedback for robots to learn tasks such as product handling, warehouse operations, and home services.

图片

To transform scenario resources into trainable data, JD is advancing the construction of an Embodied Intelligence Data Collection Center, planning to cumulatively collect over 10 million hours of high-quality real-world human scenario video data and over 1 million hours of robot hardware operation data within two years. The relevant collection will be strictly conducted in accordance with laws and regulations, ensuring legality, compliance, safety, and trustworthiness. Currently, it has covered over 100 specific production and service scenarios and collected over 100,000 real-world tasks.

Around data utilization, JD Cloud has self-developed collection terminals such as JoyEgoCam and built an embodied data service system covering "collection, storage, labeling, training, evaluation, simulation, and testing." Additionally, JD has launched the JoyAI-Sim physics simulation engine, connecting human operation data, simulated environments, and robot real-world testing, compatible with data augmentation and policy validation. In completed tests, the consistency between simulation and real-world testing is approximately 90%, and real-world testing efficiency reaches 3.2 times that of physical robot testing. Tests on JD's self-developed embodied operation foundation model JoyAI-RA show that as the scale of human operation video data expands, the model's real-world task scores increase steadily.

In the computing power dimension, JD Cloud's JoyScale AI computing platform leverages computing resource pooling, heterogeneous scheduling, and software-hardware collaborative optimization to achieve unified management and high-quality scheduling of diverse computing resources. Currently, JD Cloud has partnered with domestic partners to build a fully domestic 10,000-card cluster and is planning to construct a 100,000-card cluster, providing support for large-scale model training and physical AI applications.

In the training dimension, the JoyBuilder model development platform provides end-to-end services for embodied model training, optimizing multimodal data reading and training efficiency, and connecting the closed loop of training, real-world testing, and simulation, allowing enterprises to devote more energy to model and business innovation.

In the inference and model service dimension, JD Cloud's self-developed inference engine significantly improves throughput and resource utilization efficiency in high-concurrency scenarios, reducing inference costs for typical tasks by up to 80%. The one-stop large model service platform JoyMaaS has gathered over 100 mainstream open-source models and JD's self-developed models, providing unified access, enterprise-level management, and cost governance capabilities, lowering the barrier for enterprise development and application.

On the application side, JD is promoting the integration of AI and robots into the entire supply chain process. For example, over the next five years, JD Logistics plans to procure 3 million robots, 1 million unmanned vehicles, and 100,000 drones. Zhiyuan's embodied robots have entered JD Logistics's Zhilang Warehouse, where both parties are testing the boundaries of the models and improving reliability in real-world operations, promoting robots' transition from single-point operations to process collaboration.

图片

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry