Google DeepMind updates Gemini Robotics with ER 2, introducing whole-body humanoid coordination and real-time multi-robot task planning.
Google DeepMind has released Gemini Robotics ER 2, a model that shifts humanoid control from simple upper-body movements to full-body coordination. The system now manages robots from feet to fingertips, allowing machines to walk, crouch, and stretch while manipulating objects.
In demonstrations, Apptronik's Apollo 2 robot used the model to bend over for a watering can and retrieve specific items from shelving. This update expands the physical capabilities of the hardware, enabling complex dexterity tasks such as tying trash bags, sealing Ziploc bags, or unscrewing lightbulbs using five-fingered hands.
While Google admits that movement speed still requires advancement, the shift toward whole-body coordination is a prerequisite for real-world utility. The model functions as a high-level brain, translating multimodal reasoning into physical actions through a vision-language-action (VLA) model.
Technical Architecture
Gemini Robotics ER 2 departs from traditional reasoning LLMs that pause before every single action. By integrating with the bidirectional Gemini Live API, the model plans subsequent steps while the robot is already in motion, reducing the stop-and-think latency seen in earlier versions.
According to neowin.net, the model introduces continuous video understanding. This allows the robot to monitor a task in real time, estimate progress, and detect failures. Instead of restarting a failed workflow from the beginning, ER 2 can retry specific individual steps.
Google provided benchmarks to quantify these improvements over the previous ER 1.6 version. The model achieved 91.3% accuracy in identifying the exact moment an important event occurred, with a mean error of only 0.96 seconds. It also reached 57.4% accuracy in progress-classification tests.
Beyond individual control, the update introduces support for multi-robot collaboration. This allows different types of robots to work together on shared tasks, coordinated by the same embodied reasoning framework.
Industry Context
This rollout follows a rapid iteration cycle that began in March 2025. Google released ER 1.5 in September and ER 1.6 in April, steadily refining how artificial intelligence interacts with physical environments. The transition from isolated limb control to whole-body coordination suggests a move toward general-purpose robotics rather than task-specific automation.
By decoupling the high-level reasoning (ER) from the low-level motor control (VLA), Google is building a modular stack. This architecture allows the brain to be updated without needing to rewrite the basic physics of how a robot balances or grips, accelerating the deployment of new skills.
As reported by theverge.com, the ability to handle five-fingered dexterity marks a significant leap toward human-level manipulation. The challenge now shifts from whether a robot can perform a task to whether it can do so with the speed and reliability required for commercial industrial use.
Whether these benchmarks translate to the chaos of an unplanned warehouse or home environment remains the primary unanswered question.
FAQ
What is the difference between Gemini Robotics ER and VLA?
ER acts as the high-level brain for planning and environmental understanding, while the VLA model handles the direct physical movements of the robot.
How does ER 2 improve robot movement?
It enables whole-body coordination from feet to fingertips and reduces pauses by planning the next move while the robot is already moving.
What is the accuracy of ER 2 in event detection?
Google reports 91.3% accuracy in identifying the exact moment an event occurs, with a mean error of 0.96 seconds.
Can ER 2 control different types of robots?
Yes, the model supports multi-robot collaboration across different hardware types.
