Summary:
Google's new action model advances selected manipulation tasks. Partner access, benchmark scope and physical-system validation remain distinct questions.Google DeepMind introduced Gemini Robotics 2.5 on November 10 and is rolling it out to robotics partners. The release brings updated robot-control capabilities, while access remains limited.
Who can use Robotics 2.5?
The Robotics 2.5 model page directs prospective users to an early-access waitlist. It describes a vision-language-action, or VLA, model that controls robotic hardware, including hands, grippers and humanoid bodies. Google presents whole-body movement and adaptation to changing situations as capabilities, but those descriptions do not establish compatibility with every robot or readiness for a particular workplace.
Access is easy to confuse with Gemini Robotics ER 2. That separate model became publicly available through the Gemini API and Google AI Studio in July. ER 2 handles spatial reasoning and task planning, then delegates movement to a lower-level controller. A developer can therefore experiment with the reasoning layer without having access to the new 2.5 action model.
What the task results show
These four highlighted examples come from Google's multi-step and contact-rich task chart. Higher success rates are better; bold marks the better result in each pair.
| Selected task and hardware | Robotics 2 | Robotics 2.5 |
|---|---|---|
| Close a toolbox lid · Duo | 0% | 93% |
| Move a hanging mug to a table · Duo | 29% | 75% |
| Sweep debris into a dustpan · Apollo + Sharpa | 0% | 52% |
| Pack a rice cooker · Apollo + Sharpa | 0% | 40% |
Source: Google DeepMind's developer evaluations. These selected tasks are not an independent benchmark suite or a cross-vendor ranking. The remaining failures matter as much as the gains.
Google also reports 96% success for an Apollo 2 automotive-kitting pilot with Mercedes-Benz. Trial counts and detailed success criteria are not disclosed in the announcement.
Safety tests measure a different question
The accompanying ASIMOV-Jailbreak documentation describes 845 adversarial episodes across 15 strategy groups, with up to ten steps. It evaluates an agent's decisions through action checks and an LLM judge. These are not 845 factory trials, and a refusal score cannot substitute for testing physical movement around people.
The distinction follows the system's architecture: a reasoning model can decide what should happen while a separate controller determines how the robot moves. Evaluating the first component does not validate the entire machine. The dataset card also requires acceptance of access conditions for its files; a public description should not be mistaken for unrestricted downloads.
What deployment teams should check next
For teams considering a pilot, the useful next step is a task-specific evaluation plan: the supported robot configuration, the definition of a successful attempt, how interventions are counted, and what happens after a dropped object or interrupted movement. Availability, integration support and failure recovery deserve separate checks before a promising demonstration becomes a production assumption.
That is the practical significance of this release: a stronger action model can expand the tasks worth testing, while access and full-system validation remain deployment decisions.
Illustration: PromptsZone, based on Google DeepMind's model-role documentation. The unchanged Gemini family logo is used for identification, not as a Robotics-specific logo or endorsement. Source: Google.
