Gemini Robotics 2 Moves From Arms to Whole Bodies: 91.3% Moment-Finding and Multi-Robot Tasks
Google DeepMind released Gemini Robotics 2 on July 30, moving its physical AI stack beyond tabletop manipulation and upper-body humanoid control. The new system includes three models: Gemini Robotics 2 for vision-language-action motor control, Gemini Robotics ER 2 for embodied reasoning, and Gemini Robotics On-Device 2 for local deployment on robot hardware.
The release matters because it pushes the robotics model boundary from “can manipulate objects” toward “can coordinate bodies, tools, and time.” That is where most physical AI demos still break.
Three Models, Three Jobs
| Model | Role | New Capability |
|---|---|---|
| Gemini Robotics 2 | VLA motor control | Full humanoid control from feet to fingertips |
| Gemini Robotics ER 2 | Embodied reasoning | Multi-step task planning, video progress tracking, multi-robot coordination |
| Gemini Robotics On-Device 2 | Local VLA | Adapts to new robot embodiments with a few hours of data |
Gemini Robotics 2 can control Apptronik’s Apollo 2 humanoid across whole-body tasks such as walking to a shelf, picking up a watering can, and placing it in a lower bin. The same checkpoint is shown across Apollo 2 variants and a Franka Duo setup, which is the real claim: transfer across robot bodies without rebuilding the whole policy.
DeepMind says the model now handles five-finger dexterity through a 22-degree-of-freedom hand, including tasks such as tying knots, sealing ziplock bags, and manipulating objects with parallel grippers. Those are small actions, but they are exactly the actions that expose whether a robot has robust contact control or just a good demo script.
ER 2 Is the Agent Layer
Gemini Robotics ER 2 is the high-level planner. It watches video, understands instructions, calls tools, tracks whether a task is progressing, and hands motor execution to lower-level VLA models or robot APIs. Developers can now access ER 2 through the Gemini API and Google AI Studio, with a private preview on the Gemini Enterprise Agent Platform.
Google also added two developer endpoints:
| Endpoint | Use Case |
|---|---|
gemini-robotics-er-2-preview | Spatial reasoning, tool orchestration, progress classification, video moment finding, multi-robot coordination |
gemini-robotics-er-2-streaming-preview | Low-latency robot agents over bidirectional audio and video |
The ER 1.6 preview endpoint is scheduled to shut down on August 31, 2026, which makes ER 2 the new default path for developers building against Google’s robotics APIs.
The Number to Watch: 91.3%
The most useful benchmark in the announcement is not a humanoid walking clip. It is moment-finding.
Gemini Robotics ER 2 reaches 91.3% accuracy on moment-finding tasks with a 0.96-second mean absolute distance, while running at 4x the execution speed of larger model categories. Moment-finding is the ability to identify the exact video frame where a critical event occurs: when to stop pouring coffee, when a light bulb is tight, when a trash bag is tied, when a step has actually completed.
For physical robots, this is not a cosmetic metric. A robot that cannot tell whether a step is done either stops too early, repeats unnecessary motions, or ploughs ahead into a failed state. Long-horizon robotics needs progress awareness more than it needs another polished pick-and-place clip.
ER 2 also scores 57.4% accuracy on continuous progress classification, where each video frame is assigned to one of five progress bands from 0-20% through 80-100%. That is still imperfect. It is also the right shape of benchmark for robots that must recover mid-task rather than restart from scratch.
Multi-Robot Coordination
Gemini Robotics 2 adds multi-robot collaboration, allowing different machines to share a semantic plan and divide work. DeepMind demos Apollo 2 and Franka F3 Duo collaborating, and separately shows ER 2 orchestrating Boston Dynamics Spot through navigation and manipulator APIs.
The market implication is obvious: Google wants the model layer, not the robot body. If ER 2 can sit above humanoids, wheeled robots, arms, and custom industrial systems, DeepMind becomes the orchestration layer for physical AI in the same way frontier labs are trying to become the orchestration layer for digital agents.
That also explains the on-device model. Many factories, warehouses, hospitals, and defence-adjacent sites cannot tolerate cloud dependency for every action. A local VLA that adapts to a new robot body with a few hours of data is the edge version of the same platform strategy.
What Is Still Missing
The release gives meaningful progress metrics, but not enough economics. There is no per-hour operating cost, no failure-rate comparison across multi-hour tasks, no public benchmark against Physical Intelligence, Skild, Generalist AI, or robotics-specific open models. The whole-body clips are useful, but they do not yet answer whether the system can run safely for thousands of task-hours without expert reset.
Still, the release changes the competitive map. Google is now shipping a developer-accessible embodied-reasoning model with video, audio, function calling, progress tracking, and multi-robot coordination. In robotics, the software platform race just became much more explicit.