Google DeepMind has released Gemini Robotics ER 2, a new artificial intelligence model designed to serve as the "high-level brain" for robots, enabling them to reason about their surroundings in real time, plan complex multi-step tasks, and collaborate with other robots to complete workflows that a single machine couldn't handle alone. The model is now available to developers through the Gemini API, Google AI Studio, and in private preview on the Gemini Enterprise Agent Platform.

What Makes This Different From Previous Robot AI Models?

Gemini Robotics ER 2 represents a significant upgrade over its predecessor, Gemini Robotics ER 1.6. The key breakthrough is that robots can now watch continuous video feeds to track their own progress, adapt when something goes wrong, and know exactly when to move on to the next step. This means robots can self-correct in real time rather than requiring human intervention or restarting entire workflows when a task doesn't go as planned.

The model achieves this through two foundational capabilities. First, "progress classification" allows robots to track how far along they are toward completing a task by analyzing video frames and assigning them to five levels of progress ranging from 0 to 100 percent. Gemini Robotics ER 2 achieves 57.4% accuracy on these progress classification tasks, outperforming previous generation models. Second, "moment-finding" enables robots to identify the exact video frame where a critical event takes place, such as when to stop pouring coffee into a cup. The model achieves 91.3% accuracy on moment-finding tasks with a mean absolute distance of 0.96 seconds, delivering the sub-second latency actually required to safely operate physical robotics in the real world.

How Can Developers Use Gemini Robotics ER 2 to Build Better Robots?

• Real-Time Video Understanding: Developers can stream multimodal video, audio, or text directly into the model, allowing robots to see and understand their environment continuously rather than relying on static snapshots or pre-programmed instructions.
• Tool Orchestration: Developers can declare low-level control interfaces like Vision-Language-Action models or navigation APIs as tools, and the ER 2 model orchestrates these tools to complete multi-step tasks while the robot simultaneously performs actions without jarring "stop-and-think" pauses.
• Multi-Robot Collaboration: Different types of robots can now communicate via shared semantic understanding to hand off and complete complex tasks that would be impossible for a single machine, such as one robot fetching an item while another arranges it in a specific location.

To illustrate the practical application, Google DeepMind built a demonstration with Spot, a robot from partner Boston Dynamics. Gemini Robotics ER 2 orchestrates Spot's APIs for navigation and manipulator movement, creating an interactive robot that can fetch objects based on natural language commands.

What About Safety and Real-World Deployment?

Safety is a critical concern when deploying robots in human environments. Gemini Robotics ER 2 achieves significant gains on Safety Instruction Following and Human Proximity benchmarks, which evaluate how well the model adheres to physical constraints during reasoning tasks and maintains spatial awareness for detecting humans. In testing, the model successfully halted a humanoid robot when a person was nearby and autonomously resumed work only once the area was clear.

Google DeepMind is introducing a new benchmark called ASIMOV-Agentic that evaluates a foundation model's ability to act as a safe Vision-Language-Action orchestrator by testing its capacity to enforce safety constraints, monitor the environment, assess physical feasibility, and seek human clarification when needed.

How Does This Fit Into the Broader Gemini Robotics 2 Ecosystem?

Gemini Robotics ER 2 is one component of a larger system called Gemini Robotics 2, which includes three highly capable models working together. The first is Gemini Robotics 2 itself, the most advanced vision-language-action model that converts vision and language input into motor control, enabling robots to take action. This model can control full humanoids from feet to fingertips and other bi-arm robots, bringing a new level of dexterous manipulation on both hands and grippers.

The second component is Gemini Robotics On-Device 2, the most efficient vision-language-action model optimized to run locally on robotic devices without requiring internet connectivity. This model can achieve fast adaptation to completely new robot embodiments with just a few hours of data, typically with fewer than 200 examples.

Together, these models enable robots to reason through every movement and unlock a broad range of tasks. For example, a humanoid robot can walk, crouch, stretch, and manipulate objects to clean up a cluttered room, and it can even team up with other robots to finish the job faster. The profound intelligence can also run locally on-device while seamlessly adapting to entirely new robotic bodies in just a few hours.

What Does This Mean for the Physical AI Industry?

The release of Gemini Robotics ER 2 signals that the physical AI industry is moving beyond narrow, pre-programmed tasks toward more adaptable and intelligent systems. Most robots today are either pre-programmed for repetitive sequences or require teleoperation by humans. They lack the ability to truly learn for themselves or adapt to unpredictable environments. Transferring learned skills from one robot body to another has remained incredibly difficult.

By providing robots with real-time reasoning capabilities, video understanding, and the ability to collaborate with other machines, Google DeepMind is addressing some of the fundamental challenges that have limited robot deployment in real-world environments. The fact that the model can operate at sub-second latency while maintaining safety constraints suggests that practical deployment in human-shared spaces may be moving closer to reality.

The availability of these models through public APIs and developer platforms also democratizes access to advanced robotics AI, allowing companies beyond Google to build and deploy physical AI systems. This could accelerate innovation across industries ranging from manufacturing and logistics to healthcare and domestic assistance.