Gemini Robotics ER 2 Launches in Public Preview
Gemini CLIGoogle made Gemini Robotics ER 2, its latest embodied-reasoning model for robotics, publicly available through the Gemini API and Google AI Studio. The model, built on Gemini 3.5 Flash, adds whole-body control, temporal intelligence for tracking task progress, and multi-robot collaboration, letting different machines coordinate through shared semantic understanding. Alongside the launch, Google announced that the older gemini-robotics-er-1.6-preview endpoint will be shut down on August 31, 2026.
Key Takeaways
- Gemini Robotics ER 2 is now publicly available through the Gemini API and Google AI Studio, exposing two new endpoints:
gemini-robotics-er-2-previewandgemini-robotics-er-2-streaming-preview. - The model is built on Gemini 3.5 Flash and functions as a high-level "brain" that interprets natural language, plans multi-step tasks, and hands off motor execution to lower-level vision-language-action models.
- A new temporal intelligence capability lets robots continuously monitor task progress and confirm when an action, such as unscrewing a light bulb, has actually been completed — achieving a reported 92% success rate in testing.
- Multi-robot collaboration allows different robot types, like a wheeled rover and a humanoid, to share semantic understanding and divide labor across a shared task.
- Google demonstrated the model controlling Boston Dynamics' Spot and an Apptronik Apollo 2 humanoid, showing real-world navigation and object manipulation driven entirely by natural-language commands.
- The older
gemini-robotics-er-1.6-previewendpoint is being retired on August 31, 2026, giving developers roughly a month to migrate existing robotics integrations to ER 2.
Sources & Mentions
5 external resources covering this update
Gemini Robotics 2 brings whole body intelligence to robots
Hacker News
Google DeepMind Ships Three Physical AI Models For Whole Body Control, Dexterity And Multi Robot Collaboration
MarkTechPost
Gemini Robotics 2 Controls Full Humanoids: Legs, Torso, Arms, and Fingers Under One Policy
TechTimes
Google Expands Its Robotics AI
Yahoo Finance
Google DeepMind's Gemini Robotics 2 controls whole humanoids
The Next Web
Google Opens Up Gemini Robotics ER 2 to Developers
Google introduced Gemini Robotics ER 2, the newest version of its embodied-reasoning (ER) model family, making it publicly accessible through the Gemini API and Google AI Studio. Unlike Gemini's general-purpose chat and coding models, ER models are purpose-built vision-language models that act as a "high-level brain" for robots — interpreting natural language, perceiving the physical world, and planning multi-step tasks, while handing off actual motor execution to lower-level vision-language-action (VLA) models running on the robot itself.
Gemini Robotics ER 2 is built on top of Gemini 3.5 Flash and accepts interleaved text, image, video, and audio input, with a context window of up to 128K tokens and output of up to 64K tokens. Two new endpoints are now available: gemini-robotics-er-2-preview for standard request/response use, and gemini-robotics-er-2-streaming-preview, which connects through the Live API for real-time, bidirectional streaming so a robot can keep reasoning about upcoming steps without pausing execution.
Whole-Body Control and Temporal Intelligence
A major addition in this release is whole-body control, extending the model's reasoning beyond arm and gripper manipulation to an entire robot's body — letting it walk across a room, crouch to pick something off the floor, or reach up to a shelf under a single unified policy. Google paired this with temporal intelligence, which allows a robot to continuously monitor its own progress and confirm when a task has genuinely been completed rather than just executing a fixed action sequence. In testing, Google reported a 92% success rate on a fine-motor benchmark task (unscrewing a light bulb), which requires sustained full-body coordination and self-correction.
Multi-Robot Collaboration
Gemini Robotics ER 2 also introduces multi-robot collaboration, enabling different types of robots — for example a wheeled rover and a humanoid — to communicate through a shared semantic understanding of a task and divide up work that would be difficult for a single machine to handle alone. Google demonstrated this alongside partner hardware, including Boston Dynamics' Spot (controlled via natural-language navigation and manipulator commands) and an Apptronik Apollo 2 humanoid, which was shown walking across a room, picking up a watering can, and placing it on a specified shelf entirely from a spoken instruction.
Availability and Deprecation
The new endpoints accept text, image, video, and audio inputs and support function calling with blocking behavior for physical robot actions, an important safety consideration when a model's output can trigger real-world movement. Access is available today via the Gemini API and Google AI Studio, with a private preview also rolling out on the Gemini Enterprise Agent Platform. Google's VLA and on-device robotics models remain gated to early partners and were not part of this public release.
Alongside the launch, Google confirmed that the previous-generation gemini-robotics-er-1.6-preview endpoint will be shut down on August 31, 2026, giving developers building on the older model roughly a month to migrate their integrations to ER 2.