Google introduces Gemini Robotics ER 2, a high-level brain for robots that lets them reason in real time, talk with humans, and plan multi-step tasks. What’s the idea? That robots don’t just understand space, but think fast and act without awkward pauses between thought and movement.
What is Gemini Robotics ER 2
Gemini Robotics ER 2 is an "embodied reasoning" model designed to orchestrate robots’ work in real environments. Think of it like a conductor: it makes high-level decisions, coordinates sensors and low-latency models, and delegates motor execution to specialized systems like Vision-Language-Action (VLA) models or navigation APIs.
It’s not just theory. Google showed practical demos, for example using Spot from Boston Dynamics to fetch a popcorn after a natural language instruction. Those examples are available on GitHub if you want to try them.
Key improvements over the previous version
Gemini Robotics ER 2 improves critical aspects for operating in the physical world:
- Observes continuous video to follow task progress and correct mistakes without restarting.
- Coordinates multiple robots in the same space to complete complex workflows.
- Integrates with
Gemini Live APIand offers a bidirectional streaming endpoint optimized for low-latency tasks.
This avoids that annoying "stop to think" effect: the model can plan the next steps while the robot is already acting.
Temporal intelligence and progress tracking
One key question in robotics is: how does a robot know a task is really finished? Gemini Robotics ER 2 introduces two practical capabilities:
-
Progress classification: it assigns each video frame to progress levels (0-20%, 20-40%, etc.). In evaluations it reaches 57.4% accuracy on this task, giving situational awareness to adjust or repeat steps.
-
Moment-finding: it locates the exact frame where a critical event happens, for example the moment you stop seeing liquid while pouring. Here it reaches 91.3% accuracy and 0.96 seconds mean absolute distance, running four times faster than previous generations. That means decisions in under a second — essential for safe operation.
Collaboration between robots
Not all robots are the same. A wheel moves better on smooth floors, an articulated arm shines on uneven ground. ER 2 lets different robots share a semantic understanding and hand off tasks. Google shows machines from Apptronik and Franka working together to solve jobs a single robot couldn’t complete.
Improved spatial intelligence
ER 2 also strengthens spatial reasoning in three areas:
- Success or failure detection using continuous video to catch incidents in real time.
- Generalized instrument reading: from circular dials to digital screens and liquid thermometers.
- Better spatial Visual Question Answering for questions about the scene.
Together, the model achieves the top scores on several internal Google benchmarks for success detection, instrument reading, and visual questions.
Safety and physical constraints
Safety is central. ER 2 improves following safety instructions and detecting human proximity. In tests, it stops a humanoid robot when a person approaches and only resumes the task when the area is safe. Google also introduces a new benchmark to evaluate a coordination model’s ability to apply physical constraints, monitor the environment, and ask humans for clarification when needed.
Availability and how to get started
Gemini Robotics ER 2 is available to developers through the Gemini API, Google AI Studio, and in private preview on the Gemini Enterprise Agent Platform. Google publishes examples and configurations to help integrate the model with VLAs, navigation APIs, and teleoperation systems.
If you work in robotics or want to experiment, this release gives you an entry point to build faster, more reliable physical agents.
What this means for the near future
This isn’t a distant promise: this generation shows practical improvements in timing, safety, and machine collaboration. Robots that understand when a step is complete, correct errors in real time, and work together open real possibilities in logistics, light manufacturing, inspection, and personal assistance.
Does this mean everything is automated tomorrow? Not exactly. But it does lower important technical barriers and makes it easier for multidisciplinary teams to build useful robotic solutions today.
Original source
https://blog.google/innovation-and-ai/models-and-research/google-deepmind/gemini-robotics-er-2