谷歌DeepMind与波士顿动力在CES 2026宣布合作，将Gemini Robotics模型部署到包括人形机器人Atlas和机器狗Spot在内的多种机器人上，并计划在未来数月内于现代汽车的工厂进行测试。这一部署标志着人形机器人从展示性动作（如跳舞和翻滚）向实际体力劳动过渡的关键一步，目标是让机器人在陌生环境中识别并操作物体。尽管尚未披露具体性能指标，但该合作被视为制造业中通用人形机器人大规模应用的早期试验。
从产业结构看，竞争规模已呈数量化扩张。美国已有十余家公司开发人形机器人，包括Agility Robotics、Figure AI和Tesla等；在海外，竞争更为激烈，据中国行业协会CMRA统计，约有200家中国企业正在研发人形系统。波士顿动力本身经历多次资本变迁：2013年被谷歌收购，2017年转售软银，2021年由现代获得控股权。这一背景使其具备在真实工业环境中部署新一代AI机器人的资源与场景优势。
技术趋势显示，人工智能正从纯软件领域向物理世界快速延伸。Gemini被设计为多模态模型，可处理视觉、动作与环境信息，目标是提升机器人对现实世界的理解能力。与此同时，工业部署也引入新的风险维度，需要在现有安全系统基础上叠加AI推理机制以预防危险行为。总体来看，该合作反映出两个并行趋势：一是人形机器人开发主体数量的快速增长，二是通过真实工厂数据形成AI—机器人反馈闭环，以加速物理智能的规模化演进。
Google DeepMind and Boston Dynamics announced a CES 2026 partnership to deploy the Gemini Robotics model across multiple robots, including the humanoid Atlas and the quadruped Spot, with factory tests at Hyundai facilities planned within the next few months. The move signals a shift from demonstration-focused abilities, such as dancing and acrobatics, toward practical manual labor, enabling robots to navigate unfamiliar environments and manipulate objects. While no quantitative performance metrics were disclosed, the deployment represents an early industrial trial of general-purpose humanoids in manufacturing settings.
Industry competition has expanded rapidly in numerical terms. More than a dozen US companies are developing humanoid robots, including Agility Robotics, Figure AI, and Tesla, while overseas competition is larger, with roughly 200 Chinese firms working on humanoid systems, according to the CMRA industry association. Boston Dynamics itself has undergone multiple ownership changes—acquired by Google in 2013, sold to SoftBank in 2017, and taken under Hyundai’s controlling stake in 2021—positioning it to combine industrial scale with advanced AI integration.
The broader trend shows AI moving decisively into the physical world. Gemini’s multimodal design is intended to process vision, action, and environmental context, improving robots’ physical intelligence. However, embedding AI in industrial machines also introduces safety risks, requiring layered safeguards and predictive reasoning to prevent harmful behavior. Overall, the collaboration highlights two converging dynamics: rapid growth in the number of humanoid developers and the use of real factory deployments to create data-driven feedback loops that accelerate scalable physical intelligence.