Gemini Robotics 2 Brings Whole Body Intelligence to Robots
Source: https://deepmind.google/blog/gemini-robotics-2-brings-whole-body-intelligence-to-robots
Author: Carolina Parada, Google DeepMind
Published: 2026-07-30
Overview
Google DeepMind introduced Gemini Robotics 2, the intelligence layer for the next generation of adaptable robots. This release unlocks intelligent whole-body control, advanced dexterity, and multi-robot collaboration — a major step toward general-purpose physical AI.
Three Models
1. Gemini Robotics 2 (VLA): Most advanced vision-language-action model converting vision+language into motor control. Capable of controlling full humanoids from feet to fingertips and bi-arm robots. New dexterous manipulation on both hands and grippers.
2. Gemini Robotics ER 2 (VLM): Most capable embodied reasoning model. Acts as the robot's high-level brain — communicates with humans, understands the physical world, plans multi-step tasks lasting several minutes. Now supports multi-robot collaboration.
3. Gemini Robotics On-Device 2: Most efficient VLA optimized to run locally on robotic devices. Adapts to completely new robot embodiments with just a few hours of data (<200 examples).
Key Capabilities
Whole-Body Control
For the first time, the model controls entire humanoid robots (e.g., Apptronik Apollo 2): walking, crouching, stretching, and manipulating objects. The robot can navigate to a table, pick up an item, walk to shelves, and place it precisely.
Advanced Dexterity
Controls the 22-DOF five-fingered Shadowhand on Apollo 2 for delicate tasks like tying knots and sealing ziplock bags. Also operates standard grippers on Franka Duo platform for tight packing.
Multi-Robot Collaboration
Different types of robots can communicate and work together to solve workflows no single robot could do alone.
Fast Adaptation
On-Device 2 adapts to drastically different embodiments (Dexmate, SO101, Trossen) with fewer than 200 examples and a few hours of training data.
Safety
- Introduces ASIMOV-Agentic benchmark for agentic safety orchestration and uncertainty resolution
- Measures refusal of unsafe tool calls from VLA, prediction of task feasibility, and proactive human intervention requests
- Gemini Robotics ER 2 is described as DeepMind's safest robotics model to date in safety constraint following and human proximity benchmarks
Availability
- Gemini Robotics ER 2: available on Google AI Studio and in private preview on Gemini Enterprise Agent Platform
- VLA and On-Device models: available to early-access partners
中文概要
Google DeepMind 发布 Gemini Robotics 2，解锁了智能全身控制、高级灵巧性和多机器人协作。三个模型：VLA 模型可从脚到指尖控制全人形机器人（如 Apptronik Apollo 2），能结结、封密封袋；ER 2 具备多机器人协作能力，可规划持续数分钟的多步骤任务；On-Device 2 可在几小时内适应全新机器人形态（不到 200 个样本）。安全方面引入 ASIMOV-Agentic 基准，测试 Agent 拒绝不安全工具调用、预测任务可行性和主动请求人类干预。