Google launches Gemini Robotics ER 2 as a high-level reasoning layer for robots
Google DeepMind
Google announced Gemini Robotics ER 2, an embodied-reasoning model positioned as a high-level brain for robots that hands motor execution to lower-level VLA models. Google says it adds real-time spatial reasoning, multi-step planning, continuous video progress tracking, tool calls and multi-robot collaboration, and reports benchmark gains over ER 1.6. It is available via the Gemini API and Google AI Studio, with private preview on Gemini Enterprise Agent Platform.
What Google announced
Google announced Gemini Robotics ER 2, which it describes as an embodied-reasoning model acting as a high-level brain for robots. The architecture is deliberately layered: ER 2 handles reasoning, conversation and multi-step planning, then hands motor execution to a lower-level vision-language-action model. Developers declare control interfaces, such as VLA models or navigation APIs, as tools and stream video, audio or text into the model. Because it consumes continuous video rather than static snapshots, Google says robots can track their own progress, retry failed steps and know when to advance. Google also reports multi-robot collaboration, native tool calling, and benchmark results compared with ER 1.6 and other frontier models, including 57.4% accuracy on progress classification and 91.3% accuracy on moment-finding. A Boston Dynamics Spot demo illustrates orchestration of navigation and manipulator APIs.
Where a robotics engineer might start
A robotics integration engineer with an existing VLA controller could trial ER 2 as the planning layer for one bounded, multi-step task rather than as a replacement for low-level control. For example, a pick-and-place workflow that currently runs a fixed sequence could be wrapped with ER 2 receiving a camera stream and issuing tool calls to the existing VLA and navigation APIs, while a human operator watches. The conditional aim is to see whether the model's progress tracking lets the robot retry a failed step without restarting the whole sequence. Because Google reports benchmark improvements rather than production outcomes, this should be treated as an exploratory trial with human supervision, not an assumption that the published numbers transfer to a specific fleet.
Designing a measurable trial
A team could define one task with a clear success condition and a completion time, then run a baseline with the current controller alone before adding ER 2 as the orchestrator. Suggested measures: task completion rate, number of manual interventions per run, and whether failures are caught mid-execution versus at the end. If progress tracking is the hypothesis, test deliberately introduced disruptions, such as a misplaced object, and log whether the model detects the failure and retries. Include failure checks for false completion, where the model declares success while the task is unfinished, and for latency during streaming. Keep a human supervisor able to halt the robot, and review logs for safety-constraint adherence, since Google's safety claims are its own benchmark results.
Trade-offs and when not to use it
ER 2 adds a reasoning layer between intent and motion, so it introduces a dependency on streaming connectivity, tool definitions and the model's judgment about when a task is done; the reported 57.4% progress-classification accuracy suggests completion detection is imperfect and would need checking in your domain. Prerequisites include an existing lower-level VLA or API layer, a way to stream video, and safety review before letting the model orchestrate physical actions. Open questions from the excerpt include the full benchmark methodology, latency outside the described Live API setup, and the terms of enterprise availability, which Google describes only as private preview. A team should not adopt it as a safety authority or for tasks where a wrong completion judgment is costly without an independent safeguard.
Context and limitations
- Based on a short excerpt from Google's announcement; the full article and linked technical reports were not analyzed.
- Benchmark figures (e.g. 57.4% progress classification, 91.3% moment-finding accuracy) are Google-reported and not independently verified.
- Enterprise Agent Platform availability is described only as private preview; availability and terms may vary by region.
- Based on bounded source text, not independent testing. Practical suggestions are editorial proposals.