具身智能如何走出实验室、进入真实产业现场，高质量数据与持续学习闭环，是迈向规模化落地的关键。10月22至24日，QCon全球软件开发大会将在上海举行。本次大会特别策划了专题《数据驱动的具身智能：迈向规模化之路》，由蚂蚁灵波科技首席数据科学家黄用韬博士担任出品人。黄用韬博士本科毕业于清华大学数理基础科学班，博士毕业于德州大学达拉斯分校，长期从事AI基础设施与分布式系统研究与实践，在自动驾驶领域拥有多年产业经验，曾任元戎启行技术合伙人，主导构建了大规模数据与训练基础设施体系，现任蚂蚁灵波科技首席数据科学家。
How does embodied intelligence leave the laboratory and enter real industrial sites—high-quality data and continuous learning loops are the key to scaled deployment. From October 22 to 24, QCon Global Software Development Conference will be held in Shanghai. This conference specially planned the track "Data-Driven Embodied Intelligence: The Road to Scale," with Dr. Huang Yongtao, Chief Data Scientist at Ant Lingbo Technology, as producer. Dr. Huang graduated from Tsinghua University's mathematical sciences class and earned his Ph.D. from UT Dallas. He has long engaged in AI infrastructure and distributed systems research and practice, with years of industry experience in autonomous driving. He was formerly a technical partner at DeepRoute, leading construction of large-scale data and training infrastructure, and is now Chief Data Scientist at Ant Lingbo Technology.
面对真机交互数据不足、闭环效率低、模型泛化受限以及仿真与现实难以对齐等挑战，本专题将围绕具身数据体系、数据驱动模型革新、通用物理智能落地展开，探讨如何积累和利用真机数据、构建机器人通用大脑，让具身智能从实验室验证迈向规模化商用。机器人如何从真实执行中持续学习，仅有训练数据还不够，任务、环境、硬件状态、模型版本以及人工介入，都会影响对执行结果的判断。一次失败究竟需要补充数据、改进模型，还是排查硬件与标定，需要系统理解反馈后再作决策。蚂蚁灵波科技首席数据科学家黄用韬博士将分享真实世界学习系统的设计思路，通过端、边、云协同，将执行反馈、数据采集、模型迭代与效果验证衔接起来。
Facing challenges including insufficient real-machine interaction data, low loop efficiency, limited model generalization, and difficulty aligning simulation with reality, this track will unfold around embodied data systems, data-driven model innovation, and general physical intelligence deployment, exploring how to accumulate and utilize real-machine data, build robot general brains, and move embodied intelligence from laboratory validation to scaled commercialization. How do robots continuously learn from real execution—training data alone is not enough. Tasks, environments, hardware states, model versions, and human intervention all affect judgment of execution results. Whether a failure requires supplementary data, model improvement, or hardware and calibration troubleshooting requires systematic understanding of feedback before deciding. Dr. Huang Yongtao will share design ideas for real-world learning systems, connecting execution feedback, data collection, model iteration, and effect validation through device, edge, and cloud collaboration.
当机器人进入家庭、工厂和开放公共空间，如何持续获取高质量交互数据，并适应不同任务、环境与机器人本体，成为落地过程中的重要挑战。从数据采集到模型训练，再到现场部署，需要建立能够持续反馈和优化的完整链路。大晓机器人技术副总裁刘春晓博士将介绍环境式数据采集、标注与质量管理实践，解析具身基础模型如何统一建模空间理解、导航决策、操作执行及任务评估。结合晓途开放场景自主作业方案，分享跨任务、跨场景、跨本体泛化的探索，以及行业应用的部署、评测和规模化复制路径。
When robots enter homes, factories, and open public spaces, how to continuously obtain high-quality interaction data and adapt to different tasks, environments, and robot bodies becomes an important challenge in deployment. From data collection to model training to on-site deployment, a complete chain capable of continuous feedback and optimization must be established. Dr. Liu Chunxiao, Vice President of Daxiao Robotics Technology, will introduce environmental data collection, annotation, and quality management practices, analyzing how embodied foundation models uniformly model spatial understanding, navigation decisions, operation execution, and task evaluation. Combined with the Xiaotu open-scenario autonomous operation solution, he will share exploration of cross-task, cross-scenario, and cross-body generalization, as well as deployment, evaluation, and scaled replication paths for industry applications.
模型规模增长，为什么不一定带来真机表现的同步提升。随着具身基础模型不断发展，实验室效果与实际应用之间的差距仍待跨越。成功率、稳定性、泛化能力，以及工程化和成本问题，共同影响着模型能否转化为可交付的生产力。Dexmal原力灵机联合创始人汪天才将结合真机实践，分析具身基础模型的能力边界与规模化难题，探讨数据、算力和算法如何协同支撑持续迭代。圆桌对话环节，将邀请蚂蚁灵波市场运营和技术生态负责人江波、香港大学MMLab博士生陈天行、Muka Robotics联合创始人池晓威博士，围绕模型能力、数据体系、泛化挑战与实际应用展开交流。
Why does model scale growth not necessarily bring synchronous improvement in real-machine performance. As embodied foundation models continue developing, the gap between laboratory results and actual application remains to be crossed. Success rate, stability, generalization capability, and engineering and cost issues together affect whether models can convert into deliverable productivity. Wang Tiancai, co-founder of Dexmal, will combine real-machine practice to analyze embodied foundation models' capability boundaries and scaling challenges, discussing how data, compute, and algorithms collaboratively support continuous iteration. The roundtable will invite Jiang Bo, head of market operations and technology ecosystem at Ant Lingbo, Chen Tianxing, Ph.D. student at HKU MMLab, and Dr. Chi Xiaowei, co-founder of Muka Robotics, to exchange around model capability, data systems, generalization challenges, and practical applications.
这个专题的设置，和前面讨论的具身智能产业动态形成呼应。前面讨论过机器人数据采集员的工作、北京人形数据与训练基地的闭环链路、墨现科技的数采手套、弦引未来的Physical RSI、欧拉万象的Real-World RSI。这些讨论的核心，都是同一个问题，具身智能的训练数据从哪里来，怎么保证质量，怎么形成闭环。QCon把这个议题从产业新闻层面拉到了工程实践层面。黄用韬提到的端、边、云协同，刘春晓提到的环境式数据采集，汪天才提到的数据、算力和算法协同，这三个方向正好对应具身智能数据体系的三层，采集层、训练层、部署层。八道桌面任务和半小时的平均耗时，说明通用大模型离高效可靠的物理操作还有距离，而这个距离的缩短，取决于数据体系能不能支撑模型持续迭代。
The track's setup echoes the embodied intelligence industry dynamics discussed earlier. Earlier we discussed robot data collectors' work, Beijing Humanoid Data and Training Base's closed loop, Moxian Technology's data collection gloves, Xianyin Future's Physical RSI, and Ola Universe's Real-World RSI. The core of these discussions is the same question—where does embodied intelligence training data come from, how to ensure quality, and how to form a closed loop. QCon pulls this topic from industry news level to engineering practice level. Huang Yongtao's device-edge-cloud collaboration, Liu Chunxiao's environmental data collection, and Wang Tiancai's data-compute-algorithm collaboration correspond exactly to three layers of embodied intelligence data systems—collection, training, and deployment. Eight desktop tasks and half-hour average time show general large models still have distance from efficient, reliable physical operation, and shortening this distance depends on whether the data system can support continuous model iteration.
