Back to feed
News Story
APriority76
Google DeepMind (X)
4 sources

Google DeepMind Launches Gemini Robotics 2: Next-Gen Physical AI

Google DeepMind announced Gemini Robotics 2, a next-generation physical AI system enabling whole-body control for humanoid robots, advanced dexterity, and multi-robot collaboration. The system includes three models: a vision-language-action model, a real-world video understanding model, and an on-device model that adapts to new robot bodies in hours. This marks a significant step toward deploying robots in homes and workplaces.

SynthePulse Insight · AI deep reading

Gemini Robotics ER 2: A Video-Understanding-Driven Robot 'Brain'

Version 1 · 1 source

Google DeepMind releases Gemini Robotics ER 2, an 'advanced brain' for robots that leverages real-time video understanding, task orchestration, and multi-robot collaboration to enable more flexible interaction with the physical world.

  • Gemini Robotics ER 2 is an 'embodied reasoning' model that acts as a high-level brain for robots, responsible for understanding the physical world, planning multi-step tasks, and delegating execution to lower-level VLA models.
  • The model tracks progress via continuous video streams, corrects errors in real time, and knows when to proceed to the next step, showing significant improvements over ER 1.6.
  • Supports multi-robot collaboration, allowing multiple robots to work together in a shared space to complete complex workflows that a single robot cannot handle alone.
  • Now available to developers via the Gemini API, Google AI Studio, and the Gemini Enterprise Agent Platform.
  • Integrates the Gemini Live API with bidirectional streaming endpoints to optimize latency, enabling smooth task orchestration without 'stop-and-think' pauses.
  • Demonstrated in collaboration with Boston Dynamics' Spot, where the robot retrieves objects via natural language commands; code is open-sourced on GitHub.
Open section navigationFrom 'Thinking' to 'Acting': Task Orchestration Driven by Real-Time Video Understanding

From 'Thinking' to 'Acting': Task Orchestration Driven by Real-Time Video Understanding

The core breakthrough of Gemini Robotics ER 2 lies in its deep integration of video understanding with task orchestration. By continuously watching its own video stream, the model tracks task progress in real time, autonomously adjusts when errors occur, and precisely determines when to move to the next step. This 'temporal intelligence' solves a long-standing challenge in robotics: knowing when a task is complete.

Compared to the previous generation ER 1.6, ER 2 maintains a lead in tool orchestration, with evaluations covering three control modes: real VLA, simulated VLA, and human teleoperation. The model's design allows robots to 'think while acting,' simultaneously reasoning about the next step and executing the current action, avoiding the traditional 'stop-and-think' pauses.

Multi-Robot Collaboration: From Single-Unit Intelligence to Swarm Intelligence

ER 2 introduces multi-robot collaboration for the first time, enabling multiple robots to work together in a shared space to complete complex workflows that a single robot cannot accomplish independently. This marks a leap from independent task execution to swarm intelligence.

The model achieves flexible collaborative orchestration by declaring low-level control interfaces (such as VLA models or navigation APIs) as tools and streaming multimodal video, audio, or text. Developers can build multi-robot collaborative physical AI applications on this foundation.

Open Ecosystem: APIs and Open-Source Examples Lower the Development Barrier

Gemini Robotics ER 2 is available to developers via the Gemini API, Google AI Studio, and the Gemini Enterprise Agent Platform (private preview). The model integrates the Gemini Live API, using bidirectional streaming endpoints to optimize latency-sensitive tasks and enable smooth real-time interaction.

To accelerate application deployment, Google DeepMind provides configuration and prompt examples, and in collaboration with Boston Dynamics, demonstrated Spot retrieving objects via natural language commands. The relevant code is open-sourced on GitHub, allowing developers to reuse it directly.

Credibility boundary

All information in this article comes from the official Google DeepMind blog and is a first-party product launch announcement. All capability descriptions are self-reported by the company and have not been independently verified by third parties.

Insight takeaway

Gemini Robotics ER 2 provides robots with a more powerful 'brain' through video understanding, real-time task orchestration, and multi-robot collaboration, but actual performance still needs to be verified by developers through the public API.

Sources for this version

  1. Gemini Robotics ER 2: powering robotics with video understanding, task orchestration, and multi-robot collaborationGoogle DeepMind

Primary report

Google DeepMind (X)

Primary source

Same-event coverage

Also covered by 3 sources

The Verge AIT3Google DeepMind's new AI model can control a robot's entire bodyGoogle AI (X)T3Google DeepMind Unveils Gemini Robotics 2: Next-Gen AI for RobotsGoogle DeepMindT1Gemini Robotics ER 2: Powering Robotics with Video Understanding, Task Orchestration, and Multi-Robot Collaboration