Frontier Models

Google DeepMind Launches Gemini Robotics ER 2 in Public Preview

New endpoints deliver spatial reasoning and multi-robot coordination while the prior version faces deprecation on August 31, 2026.

By Marcus Vance July 30, 2026 4 MIN READ

In a spacious modern robotics laboratory with polished concrete floors and tall windows allowing soft daylight to illuminate the space a Boston Dynamics Spot quadruped robot with its distinctive yellow and black body stands alert in the foreground its articulated legs positioned precisely around scattered wooden blocks and metal pipes forming an obstacle course on the floor while two Franka F3 Duo dual robotic arms mounted side by side on a stainless steel workbench coordinate their movements one arm gripping a cylindrical metal part and the other extending to align it with millimeter accuracy demonstrating advanced spatial reasoning another Apptronik Apollo 2 humanoid robot with its sleek white and silver torso walks steadily from the left side toward the center holding a large rectangular component in its hands ready to place it onto a shared assembly platform several anonymous technicians wearing plain white lab coats and clear safety goggles stand at the edges of the room with backs or sides facing the viewer observing the ongoing multi-robot coordination task without any direct intervention the background shows additional industrial shelving units holding unlabeled equipment cases neatly arranged cables running along the walls and a large open floor area where the robots interact in a synchronized manner the Spot robot scanning the terrain ahead the Franka arms performing delicate insertions and the Apollo 2 navigating around the central workspace the entire scene captures a live action moment of multiple distinct robotic platforms operating together in a controlled real world research facility environment focused on the capabilities of the newly previewed Gemini Robotics ER 2 system the lighting is even and natural highlighting the metallic surfaces of the robots the textures of the lab coats the smooth floor reflections and the precise positioning of objects without any text logos or markings visible anywhere the composition emphasizes the integration of quadruped bipedal and stationary arm robots working in concert to complete a complex spatial task illustrating the practical application of enhanced robotic intelligence in a professional laboratory setting with every detail from the joint actuators on the Spot legs to the gripper fingers on the Franka arms and the balanced posture of the Apollo 2 conveying a truthful photojournalistic record of frontier robotics development involving entities such as Google DeepMind Boston Dynamics Apptronik and Franka in a single cohesive frame that avoids any artificial elements or compositing — Illustration: AI Intel Report

Gemini Robotics ER 2 is Google DeepMind's most capable embodied reasoning model for robotics that enhances spatial, temporal, and physical reasoning capabilities based on Gemini 3.5 Flash.

On July 30, 2026, Google released two new embodied reasoning model endpoints for robotics applications through the Gemini API.

The endpoints include gemini-robotics-er-2-preview which offers advanced spatial reasoning, agentic code execution, multi-step tool orchestration, video moment finding, progress classification, and multi-robot coordination.

The second endpoint gemini-robotics-er-2-streaming-preview is optimized for real-time text streaming using the Live API with bidirectional audio and video input.

What new endpoints support embodied reasoning in Gemini Robotics ER 2?

The gemini-robotics-er-2-preview endpoint enables robots to perform complex reasoning over visual and textual data streams in physical environments.

Developers can use the endpoint to build agents that execute multi-step tool calls while interpreting ongoing video input for task awareness.

The streaming-preview variant reduces latency for applications that require immediate responses to audio and video inputs from robot sensors.

Both endpoints extend the core Gemini model family into domains that demand physical world understanding beyond standard language processing.

How does Gemini Robotics ER 2 enable multi-robot coordination?

Gemini Robotics ER 2 supports multi-robot collaboration such as between the Apptronik Apollo 2 and the Franka F3 Duo platforms.

The model allows separate robots to share a common understanding of spatial layouts and task states during joint operations.

Coordination emerges from the model's ability to process synchronized video feeds from multiple sources and output synchronized action plans.

This feature reduces the need for custom middleware when orchestrating heterogeneous robot fleets on shared tasks.

What integration exists with Boston Dynamics Spot robot?

A demo integrates Gemini Robotics ER 2 with Boston Dynamics Spot for object fetching via natural language commands.

The Spot robot receives spoken instructions and uses the model to map commands to sequences of movements and grasping actions.

Video input from the robot's cameras feeds directly into the model for real-time progress assessment during the fetch operation.

The integration illustrates how natural language interfaces can control physical robots without intermediate programming layers.

What performance metrics does Gemini Robotics ER 2 achieve on benchmarks?

Gemini Robotics ER 2 achieves 57.4% accuracy on progress classification tasks that require assigning video frames to discrete progress levels.

For moment-finding tasks the model reaches 91.3% accuracy with a mean absolute distance of 0.96 seconds while running at four times the execution speed of larger models.

These results come at a fraction of the compute cost required by prior frontier models according to internal evaluations.

Assign video frames to 0-20% progress level
Assign video frames to 20-40% progress level
Assign video frames to 40-60% progress level
Assign video frames to 60-80% progress level
Assign video frames to 80-100% progress level

What is the foundation of Gemini Robotics ER 2?

Gemini Robotics ER 2 is based on Gemini 3.5 Flash as stated in the official model card.

The model functions as a vision-language model that adds specialized layers for spatial and temporal reasoning over physical scenes.

Continuous video feeds allow the system to maintain an internal state of task completion without requiring explicit state machines.

That’s why today we’re launching Gemini Robotics ER 2, our most capable “embodied reasoning” model for robotics.Google DeepMind

What is the timeline for deprecation of the previous model?

The gemini-robotics-er-1.6-preview model will be shut down on August 31, 2026 according to the Gemini API changelog.

Developers must migrate existing applications to the new endpoints before the shutdown to avoid service interruption.

The deprecation schedule provides a one-month window after the July 30, 2026 announcement for testing and transition.

How does this release impact the robotics industry?

The new capabilities in spatial reasoning and multi-robot coordination could accelerate deployment of collaborative robot teams in industrial settings.

Platforms from Boston Dynamics, Apptronik, and Franka may see faster integration cycles when using the standardized Gemini endpoints.

Stakeholders gain access to agentic code execution features that allow robots to generate and run their own control scripts based on observed scenes.

What are the implications for developers using the Gemini API?

Developers can access both preview endpoints immediately through the Gemini API documentation and changelog resources.

The streaming variant enables construction of low-latency agents that maintain bidirectional communication with robot hardware.

Migration from the deprecated endpoint requires updates to API calls but preserves core reasoning functions with improved performance.

Key performance metrics for Gemini Robotics ER 2 compared with prior models
Metric	Gemini Robotics ER 2	Prior Models
Progress Classification Accuracy	57.4%	Lower
Moment Finding Accuracy	91.3%	Lower
Mean Absolute Distance	0.96 seconds	Higher
Execution Speed	4x baseline	Baseline
Compute Cost	Fraction of larger models	Higher

Frequently asked

When will gemini-robotics-er-1.6-preview be shut down?

The gemini-robotics-er-1.6-preview model will be shut down on August 31, 2026.

What accuracy does Gemini Robotics ER 2 reach on progress classification?

Gemini Robotics ER 2 achieves 57.4% accuracy on progress classification tasks.

Which robots are shown in demos with the new model?

Demos include Boston Dynamics Spot for object fetching and coordination between Apptronik Apollo 2 and Franka F3 Duo.

Sources

Google — Gemini Robotics ER 2 achieves 57.4% accuracy on progress classification tasks and 91.3% accuracy with 0.96s mean absolute distance on moment-finding tasks at 4x execution speed.
Google — Two new embodied reasoning model endpoints were released on July 30, 2026: gemini-robotics-er-2-preview and gemini-robotics-er-2-streaming-preview, with gemini-robotics-er-1.6-preview scheduled for shutdown on August 31, 2026.
Google DeepMind — Gemini Robotics ER 2 is a Vision-Language-Model based on Gemini 3.5 Flash that enhances spatial, temporal, and physical reasoning capabilities.