Google DeepMind released Gemini Robotics ER 2 on July 30, 2026, a vision-language model designed to act as a robot's "high-level brain" that watches continuous video, reasons about physical tasks, and coordinates multi-step actions in real time. The model achieves 91.3% accuracy at identifying critical moments in tasks, like knowing exactly when to stop pouring coffee, while running four times faster than its predecessor. This represents a fundamental shift in how robots understand and execute complex work, moving beyond analyzing static images to reasoning over continuous video streams.

What Makes Gemini Robotics ER 2 Different From Earlier Versions?

The jump from Gemini Robotics ER 1.6 (released in April 2026) to ER 2 is more significant than the version number suggests. Earlier models reasoned over static images or short video clips, making decisions in isolated moments. Gemini Robotics ER 2 fundamentally changes this approach by processing continuous video streams and tracking task progress across minutes rather than seconds.

Think of the model less as a robot controller and more as a project manager standing next to the robot. It takes in continuous video, audio, and text, reasons about the physical scene, plans multi-step tasks, and issues commands to lower-level systems that handle actual movement. Because it runs on Google's Gemini Live API's bidirectional streaming endpoint, the model can think about the next step while the current one is still executing, eliminating the stop-and-think pauses that plagued earlier robotics agents.

What Are the Three Core Capabilities That Set This Release Apart?

• Continuous Progress Classification: The model watches video feeds and categorizes task progress into five bands (0-20%, 20-40%, 40-60%, 60-80%, and 80-100%), frame by frame. Google reports 57.4% accuracy on this metric, which outperforms both its prior generation and competing frontier models. This capability allows robots to retry a single failed step instead of restarting an entire workflow.
• Precision Moment-Finding: The model identifies the exact frame where something critical happens, enabling robots to know when to stop, switch tasks, or flag a problem. Gemini Robotics ER 2 achieves 91.3% accuracy with a mean absolute distance of 0.96 seconds, landing close to models several times its size while running four times faster than ER 1.6.
• Multi-Robot Collaboration: This is new to the ER line. Gemini Robotics ER 2 can coordinate two different robot bodies on a shared task through common semantic understanding rather than hardcoded handoff logic. Google's demonstration pairs Apptronik's Apollo 2 humanoid with a Franka F3 Duo arm to show this capability in action.

These three capabilities represent the actual structural delta between versions, which is why DeepMind paired the release with a name change rather than calling it ER 1.7.

How Does Safety Factor Into This New Model?

Google says Gemini Robotics ER 2 is its safest robotics model to date, built on improvements in Safety Instruction Following and Human Proximity benchmarks. In practical terms, the model can halt a humanoid robot when it detects a person nearby and resume only once the area clears. The company also introduced a new benchmark called ASIMOV-Agentic, which tests whether the model refuses unsafe tool calls from a vision-language-action (VLA) model and knows when to ask a human for clarification instead of guessing.

This distinction matters more than it might initially appear. A model that is merely accurate at spatial reasoning can still be dangerous if it does not know the limits of its own certainty. ASIMOV-Agentic explicitly tests for that gap, not just raw task accuracy, which represents a more mature approach to robotics safety.

What Does the Full Gemini Robotics Lineup Look Like?

Gemini Robotics ER 2 is the most openly accessible model in the entire Gemini Robotics family right now. It is the only embodied reasoning model available via general developer access through the Gemini API rather than gated behind a partner waitlist. The broader family includes several complementary models released over the past 16 months:

• Gemini Robotics (March 2025): The first model to output physical actions as a Gemini modality, available to trusted testers.
• Gemini Robotics ER (March 2025): Provides spatial understanding for roboticist-built programs, also available to trusted testers.
• Gemini Robotics On-Device (June 2025): A vision-language-action model that runs locally without internet connection, available to early-access partners.
• Gemini Robotics ER 1.6 (April 2026): Added agentic vision and instrument reading capabilities, now available via the Gemini API and AI Studio.
• Gemini Robotics 2 (July 30, 2026): A full-body vision-language-action controller for whole humanoid control, available to early-access partners.
• Gemini Robotics On-Device 2 (July 30, 2026): An offline version with new-embodiment adaptation in hours using under 200 examples, available to early-access partners.

A clear pattern emerges from this timeline: DeepMind ships the embodied-reasoning brain slightly ahead of or alongside the motor-control body, then widens access to the reasoning model first. If this pattern holds, expect Gemini Robotics 2 and On-Device 2 to move from partner waitlist to general access before the end of 2026, mirroring how ER 1.6 preceded today's wider release by roughly three and a half months.

What Are the Practical Limitations to Keep in Mind?

The 57.4% accuracy score on progress classification is not a number Google would want running unsupervised on a factory line yet. Google frames this as outperforming the field, and based on publicly benchmarked results, that claim holds true. However, there is a meaningful difference between "best available" and "reliable enough to trust alone," and that gap is exactly what robotics coverage will focus on over the next 12 months.

The release signals that DeepMind is treating July 30, 2026 as a coordinated platform release rather than a single model drop. Both the embodied reasoning line and the vision-language-action line synchronized their release schedules for the first time since March 2025, suggesting Google views these models as complementary pieces of a larger robotics ecosystem rather than standalone tools.

" }