Google DeepMind on July 30 announced Gemini Robotics 2, a new intelligence layer for next-generation robots. The company said the release adds whole-body control, advanced dexterous manipulation, and multi-robot collaboration, pushing robots beyond pre-programmed single-purpose automation toward more general AI systems that can learn and solve complex problems in the physical world.
DeepMind presented the work under the title “Gemini Robotics 2 brings whole-body intelligence to robots.” According to the company, the system is designed to let robots adapt to unpredictable environments and transfer learned skills quickly across different robot embodiments.
Three models split control, reasoning, and on-device adaptation
In an announcement led by DeepMind team head Carolina Parada, the company said Gemini Robotics 2 consists of three core models with different roles.
- Gemini Robotics 2: the company’s most advanced vision-language-action, or VLA, model, built to turn visual and language inputs into motor control. DeepMind said it enables full humanoid control from feet to fingertips and improves dexterous manipulation with both hands and grippers.
- Gemini Robotics ER 2: an embodied reasoning vision-language model that acts as a brain-like agent for robots. It is built to communicate with humans, understand the physical world, and plan multi-step tasks lasting several minutes. DeepMind said it also adds robot team collaboration.
- Gemini Robotics On-Device 2: an on-device VLA model aimed at high adaptation efficiency. DeepMind said it can usually adapt to a new robot body within hours using fewer than 200 examples.
Humanoid robots can translate intent into full-body action
In its capability demonstrations, DeepMind said Gemini Robotics 2 can translate intent into intelligent whole-body motion for humanoid robots. In one example, Apptronik’s Apollo 2 humanoid robot received the instruction to place a watering can into a green box on the bottom shelf. The robot then walked to a table, picked up the watering can, moved to the shelf, and placed it in the target box.
DeepMind also highlighted hand-level dexterity. With the 22-degree-of-freedom five-finger SharpaWave hand mounted on Apollo 2, the system was shown performing delicate tasks including knot tying and sealing zipper bags. On a standard two-finger parallel gripper setup such as the Franka Duo platform, it was also shown handling tightly packed packaging tasks.
ER 2 handles hundreds of decisions and adds a new safety benchmark
On the reasoning side, DeepMind said ER 2 can manage complex tasks involving hundreds of decisions and includes a self-correction mechanism. The model also supports multi-robot coordination, allowing different types of robots to complete workflows that would be too much for a single machine.
Safety was another focus of the release. The team introduced a new benchmark called ASIMOV-Agentic to evaluate safe coordination and uncertainty handling. DeepMind said ER 2 is currently the safest robot model in terms of following safety constraints and operating near humans.
ER 2 is live in AI Studio, while other models remain limited-access
Gemini Robotics ER 2 is now available through Google AI Studio and has entered private preview on the Gemini Enterprise Agent Platform. The VLA model and the on-device model are open only to early partners for now.
DeepMind described the launch as an important milestone toward what it called solving “physical world AGI,” adding that future robots will be able to work alongside humans on more complex challenges.