FULL STORY

Google's Gemini Robotics 2: From Launch to Testing

Google DeepMind launched Gemini Robotics 2 and ER 2, enabling full-body control and embodied reasoning. Subsequent tests show impressive robustness, though generalization remains limited.

2026-07-30 ~ 2026-08-09 · 7 episodes · 92 posts

Episode 1 · Google DeepMind Teases Gemini Robotics 2.0 (2026-07-30, 2 posts)

Google DeepMind is gearing up for a major 2.0 update to its Gemini Robotics model. Recent leaks hint at new progress and demonstrations, building anticipation following its initial April release.

Episode 2 · Google Unveils Gemini Robotics 2 and ER 2 for Full-Body Robot Intelligence (2026-07-30, 71 posts)

Google DeepMind officially released the next-generation vision-language-action (VLA) model Gemini Robotics 2 and the embodied reasoning model Gemini Robotics ER 2. These models break the limitations of traditional pre-programmed or teleoperated robots, endowing humanoid robots with full-body intelligent control, advanced dexterous manipulation, and cross-embodiment collaboration, marking a transition from tabletop tasks to full-body intelligence in real human environments.

Confirmed

  • Core release: The official launch includes Gemini Robotics 2 and ER 2, forming a model matrix driving embodied intelligence breakthroughs. ER 2 is now available on the Gemini API and Google AI Studio, and can directly connect to the Gemini Live API.
  • Full-body control: Gemini Robotics 2 coordinates actions from toes to fingertips, enabling robots to stretch, bend, and maintain balance while walking in narrow, cluttered spaces. In demos, the model provided full-body intelligence to Apptronik's Apollo 2 robot, enabling complex kitchen chores.
  • Dexterous manipulation: The model handles fine manipulation across different end effectors (e.g., robotic hands or grippers), performing tasks like tying knots and screwing in light bulbs. According to Matt Wolfe, the new model achieves precise control with 22 degrees of freedom and supports on-device operation.
  • Multi-robot collaboration: Introduces multi-robot team collaboration, allowing different robot types to communicate and coordinate in real time (e.g., handling grasping, assembly, and quality inspection), solving complex workflows that a single robot cannot complete alone.
  • Embodied reasoning: Gemini Robotics ER 2 serves as the advanced 'brain' for robots, showing significant progress in video understanding, task orchestration, advanced spatial reasoning, and multi-robot collaboration compared to the previous generation and the baseline Gemini 3.6 Flash.
  • Real-world demo: According to bousmalis, Gemini Robotics 2 completed a 20-minute uninterrupted real-time tool sorting task on FR3 Duo hardware, demonstrating generalized dexterous manipulation precision.

Why it matters

  • This release marks a step toward true general adaptability in robot intelligence. With the concept of 'one brain for any robot,' the new system significantly enhances perception, motor coordination, and multi-step task planning, promising more practical assistance in complex, unknown environments like homes and workplaces.

51 more related posts →

Episode 3 · Google DeepMind Unveils Gemini Robotics 2 Model (2026-08-02, 4 posts)

Google DeepMind has introduced Gemini Robotics 2, a new model family designed to bridge the sim-to-real gap by learning tasks entirely through simulation. The updated models also feature whole-body control, dexterous manipulation, and multi-robot collaboration.

Episode 4 · GPT-5.6 and Gemini Robotics Models Show Promise but Lack Real-World Generalization (2026-08-03, 3 posts)

Experts tested Gemini-Robotics-ER-2 and GPT-5.6-Sol, finding their high accuracy and robustness impressive but hindered by poor cross-environment generalization. Researchers emphasize that overcoming these deployment challenges, including hardware control failures, requires more real-world data.

Episode 5 · Google DeepMind Unveils Gemini Robotics 2 Model (2026-08-04, 4 posts)

Google DeepMind introduced Gemini Robotics 2, a vision-language-action model that enables full-body control of humanoid robots. Acting as a robotic brain, it analyzes live video for precise timing, supports multi-robot collaboration, and demonstrates robust generalization in unknown scenarios.

Episode 6 · Google DeepMind Unveils Gemini Robotics 2 with Whole-Body Intelligence (2026-08-07, 6 posts)

Google DeepMind officially released Gemini Robotics 2 on August 7. As a vision-language-action (VLA) model, it provides an intelligence layer for next-generation adaptive robots, combining multimodal understanding with whole-body control. The model achieves three breakthroughs: whole-body intelligent control, advanced dexterous manipulation, and multi-robot collaboration. Official videos demonstrate interaction with a robot codenamed Apollo, highlighting natural language dialogue and embodied intelligence. This marks significant progress in whole-body coordination and collaboration for embodied AI.

Confirmed

  • Model positioning: Gemini Robotics 2 is an advanced VLA model serving as the robot's intelligent brain, enabling reasoning about every action and executing complex tasks.
  • Core capabilities: Official announcements highlight three breakthroughs: whole-body intelligent control (e.g., humanoid robots walking autonomously), advanced dexterous manipulation (e.g., handling eggs), and multi-robot collaboration.
  • Official demo: Videos show interaction with a robot codenamed Apollo, focusing on natural language dialogue and embodied intelligence.
  • Behind-the-scenes: The team discussed design considerations, such as choosing between humanoid legs and wheeled robots.

Unconfirmed

  • Specific technical details, parameter scale, and training data are not disclosed in the posts.
  • Actual application scenarios and commercialization plans have not been announced.

Why it matters

Gemini Robotics 2 deeply integrates multimodal understanding with whole-body control, representing a significant advance in physical AI. Its whole-body control and multi-robot collaboration capabilities could promote the practical use of humanoid robots in complex environments, with potential impacts on industrial and service applications.

Episode 7 · Google Unveils Gemini Robotics ER 2 (2026-08-09, 2 posts)

Google DeepMind has launched the Gemini Robotics ER 2 model, providing an advanced "brain" for robots. The new model features real-time video understanding, multi-robot collaboration, and sub-second latency, enabling machines to think while moving.