Google DeepMind launched Gemini Robotics ER 2 on July 30, 2026, as its most capable embodied reasoning model for robotics, making it publicly available to developers through the Gemini API and Google AI Studio, with private preview access on Gemini Enterprise Agent Platform. The model acts as a high-level brain for robots: it can chat with humans, understand the physical world, plan multi-step tasks, call tools such as Google Search or user-defined functions, and hand off motor execution to lower-level vision-language-action models.

Gemini Robotics ER 2 upgrades Gemini Robotics ER 1.6 by using continuous video feeds to track progress, adapt when a task goes wrong, and decide when to move to the next step. Developers can declare low-level control interfaces, including Vision-Language-Action models or navigation APIs, as tools and stream multimodal video, audio, or text directly into the model. Google DeepMind said ER 2 consistently outperforms ER 1.6 for tool orchestration across real VLA, sim VLA, and human tele-op control modes.

Gemini Robotics ER 2 integrates with the Gemini Live API through a bidirectional streaming endpoint for latency-sensitive tasks, allowing robots to orchestrate multi-step actions without stop-and-think pauses. Google DeepMind demonstrated the system with Boston Dynamics’ Spot, using ER 2 to coordinate Spot APIs such as navigation and manipulator movement so the robot could fetch a popcorn snack after a natural-language command. The code is available on Github with other examples.

Google DeepMind reported progress on two video-understanding capabilities: progress classification and moment finding. In progress classification, each frame in a video feed is assigned to one of five levels: 0-20%, 20-40%, 40-60%, 60-80%, and 80-100%. Gemini Robotics ER 2 achieved 57.4% accuracy on progress classification tasks. In moment-finding, which identifies the exact video frame where a critical event occurs, ER 2 achieved 91.3% accuracy and a 0.96s mean absolute distance, while running at 4x the execution speed and a fraction of the compute cost of much larger model categories.

Gemini Robotics ER 2 also enables multi-robot collaboration, allowing machines to share semantic understanding and hand off complex tasks. Google DeepMind cited collaboration between Apptronik’s Apollo 2 and Franka F3 Duo as an example. The model also improved spatial reasoning across success/failure detection, Enhanced Spatial VQA, and general instrument reading, including tests across 10 different types of instruments.

Google DeepMind described Gemini Robotics ER 2 as its safest model, citing gains on Safety Instruction Following and Human Proximity benchmarks. The company said ER 2 halted a humanoid robot when a person was nearby and resumed work only after the area was clear. Google DeepMind also introduced a benchmark for evaluating whether a foundation model can act as a safe VLA orchestrator by enforcing safety constraints, monitoring the environment, assessing physical feasibility, and seeking human clarification.

Google DeepMind launched Gemini Robotics ER 2 on July 30, 2026, as its most capable embodied reasoning model for robotics, making it publicly available to developers through the Gemini API and Google AI Studio, with private preview access on Gemini Enterprise Agent Platform. The model acts as a high-level brain for robots: it can chat with humans, understand the physical world, plan multi-step tasks, call tools such as Google Search or user-defined functions, and hand off motor execution to lower-level vision-language-action models.

Gemini Robotics ER 2 upgrades Gemini Robotics ER 1.6 by using continuous video feeds to track progress, adapt when a task goes wrong, and decide when to move to the next step. Developers can declare low-level control interfaces, including Vision-Language-Action models or navigation APIs, as tools and stream multimodal video, audio, or text directly into the model. Google DeepMind said ER 2 consistently outperforms ER 1.6 for tool orchestration across real VLA, sim VLA, and human tele-op control modes.

Gemini Robotics ER 2 integrates with the Gemini Live API through a bidirectional streaming endpoint for latency-sensitive tasks, allowing robots to orchestrate multi-step actions without stop-and-think pauses. Google DeepMind demonstrated the system with Boston Dynamics’ Spot, using ER 2 to coordinate Spot APIs such as navigation and manipulator movement so the robot could fetch a popcorn snack after a natural-language command. The code is available on Github with other examples.

Google DeepMind reported progress on two video-understanding capabilities: progress classification and moment finding. In progress classification, each frame in a video feed is assigned to one of five levels: 0-20%, 20-40%, 40-60%, 60-80%, and 80-100%. Gemini Robotics ER 2 achieved 57.4% accuracy on progress classification tasks. In moment-finding, which identifies the exact video frame where a critical event occurs, ER 2 achieved 91.3% accuracy and a 0.96s mean absolute distance, while running at 4x the execution speed and a fraction of the compute cost of much larger model categories.

Gemini Robotics ER 2 also enables multi-robot collaboration, allowing machines to share semantic understanding and hand off complex tasks. Google DeepMind cited collaboration between Apptronik’s Apollo 2 and Franka F3 Duo as an example. The model also improved spatial reasoning across success/failure detection, Enhanced Spatial VQA, and general instrument reading, including tests across 10 different types of instruments.

Google DeepMind described Gemini Robotics ER 2 as its safest model, citing gains on Safety Instruction Following and Human Proximity benchmarks. The company said ER 2 halted a humanoid robot when a person was nearby and resumed work only after the area was clear. Google DeepMind also introduced a benchmark for evaluating whether a foundation model can act as a safe VLA orchestrator by enforcing safety constraints, monitoring the environment, assessing physical feasibility, and seeking human clarification.