Google DeepMind’s new AI model can control a robot’s entire body 谷歌DeepMind的新AI模型可控制机器人全身

TL;DR

• Gemini Robotics 2 enables whole-body control for humanoid robots, expanding from upper body to full motion including feet and fingertips.
• Enhanced dexterity allows complex manipulation tasks such as tying bags, sealing containers, and unscrewing lightbulbs using five-fingered hands.
• Gemini Robotics ER 2 improves embodied reasoning with better task duration understanding, multi-step execution, and collaborative capabilities across heterogeneous robot types.
• Safety features are significantly improved in ER 2, enabling real-time human proximity detection and automatic safe stopping mechanisms.
• The On-Device Model now supports faster adaptation to diverse robotic embodiments without requiring internet connectivity, increasing deployment flexibility.

Why It Matters

This advancement represents a critical leap toward practical, autonomous humanoid robots capable of operating safely and effectively in unstructured human environments. For AI practitioners and robotics engineers, it demonstrates progress in integrating perception, planning, control, and safety into a unified framework that scales across different hardware platforms—key for real-world deployment in homes, warehouses, and care settings.

Technical Details

• Whole-body motion control: Gemini Robotics 2 extends motor control beyond the torso to include legs, arms, and fingers, enabling coordinated actions like walking, crouching, reaching, and grasping simultaneously.
• Fine-grained hand manipulation: Support for articulated five-fingered hands allows precise object interaction, demonstrated through tasks such as securing plastic bags and handling small components.
• Embodied Reasoning (ER) upgrade: Gemini Robotics ER 2 enhances temporal awareness by recognizing task boundaries, supports long-horizon planning, and facilitates inter-robot coordination via shared command interpretation.
• Safety-aware architecture: Integrated vision-language models detect human presence dynamically and trigger predefined safety protocols, including emergency stops and reduced operational speed.
• Localized inference capability: The On-Device Model has been optimized for edge deployment, allowing rapid policy transfer to new robot morphologies—including those with unconventional kinematics or sensor suites—without cloud dependency.

Industry Insight

The integration of whole-body control, advanced dexterity, and onboard reasoning signals a shift from specialized, single-task robots to general-purpose humanoid agents adaptable across domains. Companies investing in physical AI should prioritize modular frameworks that support seamless model migration between platforms while embedding safety-by-design principles early in development cycles. Collaboration between software teams specializing in LLM-based reasoning and hardware engineers designing compliant actuators will become increasingly vital to unlock scalable autonomy in dynamic environments.

TL;DR

• Google DeepMind 发布 Gemini Robotics 2，实现从上半身控制到全身运动控制的突破，支持行走、蹲下、伸展及精细操作。
• 新增对五指灵巧手的完整控制，可完成系垃圾袋、拧灯泡等复杂任务；同时升级了具身推理模型（ER），提升多步骤任务理解与执行能力。
• 支持异构机器人协作（如 Apollo 2 指挥双臂机器人），并增强本地化部署能力，使机器人能更快适应新形态与传感器配置。

为什么值得看

该进展标志着通用型人形机器人向真实世界复杂任务迈出关键一步，尤其在全身协调、人机安全交互和跨设备协同方面具备显著工程价值，为工业与服务机器人落地提供重要技术范式。

技术解析

• Gemini Robotics 2 引入全身运动控制架构，整合足部至指尖的连续动作规划，实现动态平衡与空间操作能力，支撑弯腰取物、货架抓取等高自由度行为。
• 升级后的五手指控模块结合强化学习与模仿学习策略，在 Ziploc密封、灯泡拆卸等需要力反馈与姿态微调的任务中表现稳定，体现高精度末端执行器控制能力。
• Gemini Robotics ER 2 基于视觉语言模型构建，具备时序感知与任务边界识别能力，可理解“开始”与“结束”语义，支持长周期多步指令分解与状态追踪。
• 安全机制集成实时人体检测与紧急制动逻辑，当检测到近距离人类接近时自动触发安全停机协议，降低物理交互风险。
• On-Device Model 优化了模型压缩与迁移学习效率，可在无网环境下快速适配不同构型机器人（如改变自由度或传感器布局），提升部署灵活性。

行业启示

• 全身控制+灵巧手+具身推理的组合将成为下一代人形机器人的标准配置，推动服务与制造场景从单一动作向自主任务流演进。
• 多机器人异构协作能力的成熟将加速柔性产线与家庭辅助系统的规模化部署，未来可能出现“主从式”机器人团队模式。
• 本地化推理模型的进步意味着边缘计算在机器人系统中的权重上升，有助于降低延迟、保障隐私并提升断网环境下的可靠性。