Google DeepMind Launches Gemini Robotics ER 2, Adding Video Progress Tracking and Multi-Robot Coordination

What happened

Google DeepMind has launched Gemini Robotics ER 2, an "embodied reasoning" model that functions as a high-level planning layer for robots, orchestrating multi-step tasks, tracking progress via continuous video feeds, and coordinating between multiple robots working in shared spaces. The model is now publicly available through the Gemini API and Google AI Studio, with a private preview on the Gemini Enterprise Agent Platform.

How it works

Gemini Robotics ER 2 does not control motors directly. Instead, it hands off motor execution to lower-level vision-language-action (VLA) models while it handles reasoning, planning, tool calls (including Google Search and user-defined functions), and human interaction. Developers declare low-level control interfaces — VLA models or navigation APIs — as tools and stream multimodal video, audio, or text directly into the model.

The model integrates with the Gemini Live API through a bidirectional streaming endpoint built for latency-sensitive tasks, according to Google DeepMind, which the company says eliminates "stop-and-think" pauses during multi-step orchestration. In a demonstration built with Boston Dynamics, Gemini Robotics ER 2 orchestrated Spot's navigation and manipulator APIs to fetch objects on natural-language command; the code is published on GitHub.

Benchmark claims

Google DeepMind reports the following results for Gemini Robotics ER 2:

• Progress classification (tracking task completion across five bands from 0-20% to 80-100%): 57.4% accuracy, which the company says outperforms both the prior Gemini Robotics ER 1.6 model and competing frontier models.
• Moment-finding (identifying the exact video frame where a critical event occurs, such as when to stop pouring liquid): 91.3% accuracy with a 0.96-second mean absolute distance. Google DeepMind claims this rivals larger model categories while running at roughly 4x the execution speed and a fraction of the compute cost.
• Spatial reasoning benchmarks, including success/failure detection now operating on raw video rather than static images, general instrument reading (tested across 10 instrument types including digital displays and linear scales), and enhanced spatial visual question answering (ERQA). Google DeepMind states ER 2 achieves the highest accuracy across all three versus prior models.
• Safety benchmarks: unspecified numeric scores on "Safety Instruction Following" and "Human Proximity" tests, with the company citing a demonstration where the model halted a humanoid robot when a person approached and resumed only once the area cleared.

None of the comparison models used for these benchmarks were named in the source material, and independent verification of these figures is not yet available.

New capability: multi-robot collaboration

Gemini Robotics ER 2 introduces the ability for heterogeneous robots — for example a wheeled rover and a humanoid — to share semantic understanding of a task and hand off subtasks to each other. Google DeepMind demonstrated this using Apptronik's Apollo 2 humanoid and the Franka F3 Duo arm system working together.

Pricing for API access was not disclosed in the announcement. Context window, parameter count, and training cutoff date were also not specified.

What this means

This release positions Gemini Robotics ER 2 as a coordination and reasoning layer rather than a motion-control model, competing less with VLA specialists and more as the "brain" that sits above them — a role Google is betting becomes the bottleneck as robotics hardware diversifies across form factors. The emphasis on continuous video-based progress tracking and sub-second moment-finding addresses a real gap in current robotics stacks: knowing when a task is actually done, not just executing motion primitives blindly. The multi-robot collaboration feature is notable because it suggests Google is targeting warehouse and industrial deployments where fleets of dissimilar robots must coordinate, not just single-robot research demos. The lack of disclosed pricing, context window, and hard benchmark comparisons against named competitors makes it difficult to independently assess how much of an advance this represents versus marketing framing — claims like "4x execution speed" and "fraction of compute cost" are Google's own characterizations without a stated baseline.