Full Breakdown
Google DeepMind Unveils Gemini Robotics 2 for Whole-Body Robot Control
7/31/2026, 11:59:12 PM
Announcement and Capabilities
Google DeepMind announced Gemini Robotics 2, an AI model that can command a full humanoid robot—from feet to fingertips. Demonstrations show the Apptronik Apollo 2 robot walking, crouching, stretching, picking up a watering can, retrieving items from shelves, sealing Ziploc bags, tying trash-bag strings and unscrewing light bulbs. The model can also coordinate multiple robots, such as directing a dual-arm robot to sort tools while a humanoid cleans a garage.
Background & Context
The original Gemini Robotics release in March introduced “embodied reasoning” (ER) for arm-hand manipulation. Gemini Robotics 2 adds a Vision-Language-Action (VLA) stack that merges a vision-language model with two VLA models, enabling full-body motion planning and execution. DeepMind describes the update as a step toward “physical AGI,” where a robot could perform any task a human can.
Data & Statistics
- The VLA stack controls five-finger hands with 22 degrees of freedom.
- Benchmark success rates for whole-body tasks range from 45.7 % to 76.3 %, with multi-finger dexterity scores between 32 % and 92 % and parallel-gripper tasks between 74.2 % and 89.6 % (DeepMind data).
- The on-device variant can adapt to a new robot embodiment with fewer than 200 training examples in a few hours, allowing operation without internet connectivity.
Official Statements & Responses
Carolina Parada, head of robotics at Google DeepMind, framed the release as a milestone toward “general-purpose physical AI” that can learn and adapt in unpredictable environments.
Demis Hassabis, CEO of DeepMind, reiterated the vision of an AI operating system for robots comparable to Android for smartphones, emphasizing a “universal AI control layer” that could let hardware companies focus on mechanical design.
DeepMind also highlighted safety upgrades in the ER 2 component: the system can detect nearby humans, trigger safety tool calls and halt motion if a person approaches too closely.
Safety Considerations
Parada warned that “the safety question is even more pressing because you're putting them in a lot of other situations,” adding that “there's a lot of uncertainty that will show up, and so you want to be able to understand the safety question more deeply.” DeepMind is deploying a new benchmark, ASIMOV-Agentic, to evaluate whether robot commands could lead to harmful or uncertain outcomes.
Impact and Future Outlook
If reliability reaches industrial standards, humanoid robots could share factory floors with human workers, perform complex assembly, packaging and precision-handling tasks, and reduce the need for task-specific programming. The technology also opens possibilities for robots operating in remote or space environments where cloud connectivity is unavailable.
Conflicting Reports & Gaps
Public information does not include precise measurements of movement speed or real-world deployment metrics beyond controlled demos. Success-rate figures vary widely across tasks, indicating performance is still task-dependent.
Verbatim Quotes
- “The safety question is even more pressing because you're putting them in a lot of other situations,” — Carolina Parada, head of robotics at Google DeepMind, tells WIRED
- “There's a lot of uncertainty that will show up, and so you want to be able to understand the safety question more deeply.” — Carolina Parada, head of robotics at Google DeepMind, tells WIRED
- “Most robots are pre-programmed or teleoperated for narrow, repetitive task sequences. They lack the ability to truly learn for themselves or adapt to unpredictable environments,” — Carolina Parada, head of robotics at Google DeepMind, tells WIRED
- “Unlocking the true potential of robotics requires moving past single task automation toward general-purpose intelligence,” — Dr Parada, head of robotics at Google DeepMind, tells WIRED
