Google推出了Gemini Robotics 2系统，这是一套新的AI模型组合，使机器人能够实现整体身体控制和复杂操作1。相比前版本仅支持上半身控制，新系统已扩展到支持从脚到指尖的全身运动2，包括行走、下蹲、伸展和操纵物体等动作2。演示中，Apptronik Apollo 2机器人展示了弯腰拾取浇水壶和从架子上取下特定物品的能力2。
该系统由三个核心模型组成：Gemini Robotics 2视觉语言动作模型、Gemini Robotics ER 2具身推理模型，以及Gemini Robotics On-Device 2本地运行优化模型1。其中，On-Device 2可在少于200个示例的情况下，通常仅需几小时内就能适应新的机器人体型1。新模型能够控制具有22度自由度的五指手和两指平行夹爪，完成打结或密封拉链袋等精细操作1。
系统还新增了多机器人协作功能，支持在几分钟内执行数百个决策点的多步骤任务1。Google同步引入了ASIMOV-Agentic安全基准测试，用于评估代理安全编排和不确定性解决能力1。目前，Gemini Robotics ER 2已在Google AI Studio上提供，而VLA和On-Device模型向早期访问合作伙伴开放1。
Google has unveiled Gemini Robotics 2, an advanced AI model system that enables robots to achieve coordinated full-body control and execute complex multi-step tasks.12 The new model represents a significant expansion from its predecessor, which could only control the upper body of humanoid robots; Gemini Robotics 2 now supports whole-body motions spanning from feet to fingertips.2
The system comprises three core components: Gemini Robotics 2, a vision-language-action model for comprehensive robot control; Gemini Robotics ER 2, an embodied reasoning model; and Gemini Robotics On-Device 2, an optimized model designed to run locally on robots.1 These models enable robots to perform intricate manipulation tasks, including fine-grained dexterity work such as tying knots and sealing zippered bags using hands equipped with 22 degrees of freedom.1 Demonstrations using the Apptronik Apollo 2 robot showcase the system's capabilities, including bending to pick up watering cans and retrieving specific items from shelves.2
Beyond individual robot control, the new platform introduces multi-robot collaboration features and extends task execution to handle hundreds of decisions over periods of minutes.1 The On-Device model can adapt to new robot morphologies within hours, typically requiring fewer than 200 examples.1 Additionally, Google has introduced the ASIMOV-Agentic safety benchmark to evaluate the safety orchestration and uncertainty resolution capabilities of agent systems.1 Gemini Robotics ER 2 is now available through Google AI Studio, while the VLA and On-Device models are being provided to early access partners.1
评论
还没有评论，欢迎留下第一条。