Google DeepMind has launched Gemini Robotics 2.5, a vision-language-action model designed to convert visual and linguistic inputs into precise motor control. While DeepMind reports 20 to 30 percent faster execution and significant safety improvements, this remains a partner-only release rather than a public API. A successful proof-of-concept at Mercedes-Benz factories demonstrated a 96 percent success rate in automotive kitting, yet the model relies on cloud-based processing. For developers, the lack of public access means real-world integration depends on specific hardware compatibility, network latency, and the ability to manage recovery from the remaining four percent failure rate.
Google DeepMind introduced Gemini Robotics 2.5 on November 10, 2026, reporting higher task success rates, faster execution, and improved physical-safety behavior. The company also described a production-line proof of concept in which Apptronik’s Apollo 2 humanoid robots achieved a 96% success rate while sorting automotive parts at Mercedes-Benz factories.
The availability detail is straightforward: Gemini Robotics 2.5 is rolling out to DeepMind’s robotics partners. The announcement introduces a named vision-language-action model, but it does not announce a general developer API or public model download. The reported performance gains therefore arrive ahead of broad access for outside evaluation. Google DeepMind’s announcement supports both the technical claims and that narrower release status.
What Has Launched, and Who Can Use It?
Gemini Robotics 2.5 converts visual and language inputs into motor control. In practical terms, it connects what a robot sees and what it has been asked to do with movements that carry out the task.
Gemini Robotics 2.5: The vision-language-action, or VLA, model that generates robot actions.
Gemini Robotics ER 2: An embodied reasoning model for understanding physical spaces and planning complex, multistep work.
Gemini Robotics On-Device 2: A lightweight VLA model optimized to run locally on robotic hardware.
Those names matter when assessing access. The page’s invitation to try Gemini Robotics ER 2 refers to the reasoning model, not public access to Robotics 2.5’s motor-control capabilities. Likewise, the existence of an on-device model does not mean this 2.5 release runs locally.
The model page lists research partners including Apptronik, Agile Robots, and Boston Dynamics. It also describes a program involving more than 100 trusted testers and offers an early-access waitlist. That broader program should not be read as confirmation that every tester already has Robotics 2.5.
For developers, I’d treat this as a partner model launch, not an API launch that can immediately become a weekend robot project. The next steps depend on access, compatible hardware, integration requirements, and deployment conditions, none of which the announcement packages into a public setup guide.
What Google’s Performance Numbers Actually Say
Like this article (0 likes)
Save article
DeepMind says it tested Gemini Robotics 2.5 across more than 200 distinct tasks. Examples include serving food with tongs, tying a bow on a gift box, and pipetting liquid. These activities exercise different kinds of manipulation: tool use, flexible materials, and delicate objects.
The company reports the following results:
Measure
DeepMind’s reported result
Scope or condition
Evaluation breadth
More than 200 distinct tasks
Multiple kinds of manipulation
Task success
30% to 100% higher success rate
Varies by task type, compared with previous baselines
Task execution
Tasks completed 20% to 30% faster
Compared with previous baselines
Factory kitting
96% success rate
Mercedes-Benz production-line proof of concept
Safe task completion
Improvement of 33% around obstacles and 43% around nearby people
DeepMind’s safety evaluations
These are vendor-reported measurements. The opened coverage does not supply independent replication, and the announcement does not provide enough task-level detail to reconstruct the headline comparisons. DeepMind describes the comparison as being against previous baselines, so it would be too specific to label every result a direct comparison with one earlier released model.
The success-rate wording also needs care. A 30% higher success rate is not a gain of 30 percentage points. As an illustration, a baseline success rate of 50% would become 65% with a 30% relative improvement. That example explains the arithmetic; it is not a published Robotics 2.5 result.
The task examples offer useful context beyond the percentages. A ribbon changes shape as a robot manipulates it, while tongs require control through a tool rather than direct contact with the object. A model that improves across these different conditions is addressing a broader problem than repeating one fixed movement.
DeepMind says Robotics 2.5 works with both Apollo 2’s five-finger, 22-degree-of-freedom SharpaWave hand and the Franka Duo platform’s two-finger parallel grippers. That hardware range is meaningful, although the aggregate scores do not tell readers whether each robot achieved comparable improvements on the same tasks.
Mercedes-Benz Provides a Factory Test, Not a Deployment Verdict
Apptronik’s Apollo 2 robots performed automotive kitting on active Mercedes-Benz production lines in Kecskemét, Hungary, and Sindelfingen, Germany, according to DeepMind.
Kitting means taking parts from bulk supplies and arranging them into trays or containers in the sequence needed for assembly. In this proof of concept, Apollo 2 picked up metal automotive brackets from supply bins and placed them into sequencing containers, correcting gaps as it worked.
The setup used two model roles. Gemini Robotics ER tracked the workflow, while Gemini Robotics 2.5 controlled the hands and arms. DeepMind’s account names the ER family but does not specify the exact ER version used in the pilot. The announcement reports a 96% success rate for the proof of concept.
Apptronik co-founder and CEO Jeff Cardenas described the improvement in a statement published by DeepMind:
Pairing Apollo 2 with Gemini Robotics, we saw the models improve dramatically in just a few weeks of post-training
That supplies some evidence for faster skill acquisition, but it is a partner’s description, not a controlled measurement of training time. The release’s quantified 20% to 30% speed gain concerns task execution, which is a different measure from how quickly a robot learns a skill.
The factory result is encouraging because it moves the evaluation into an operating workplace. Its practical meaning still depends on what counted as a successful attempt, how many attempts were evaluated, and how failures were recovered. Those details are not supplied in the announcement.
At face value, 96% success leaves 4% of evaluated attempts outside the success category. That arithmetic does not tell us whether an unsuccessful attempt meant a missed pickup, an incorrectly arranged part, or an intervention.
My view is that recovery is the consequential next measurement. A robot that safely detects and retries a missed pickup creates a different operational burden from one that silently delivers an incorrect kit. Ars Technica’s reporting likewise highlights the remaining reliability gap and reports DeepMind’s acknowledged tradeoff between broad generalization and very high success in a specific application.
Physical Hazard Avoidance and Jailbreak Testing Are Different
Gemini Robotics 2.5 uses inference-time reasoning guidance to steer around physical hazards, DeepMind says. Inference-time guidance operates while the model is producing its decisions, as opposed to changing the model only during training.
DeepMind reports that this improved safe task completion by 33% around obstacles and 43% around nearby people. Ars describes behaviors such as stopping or waiting until continuing is safe. These measurements concern completing tasks safely under the evaluated conditions; they are not direct measurements of workplace injury rates.
The release also introduces ASIMOV-Jailbreak, an open-source robotics safety benchmark. Ars Technica reports that it contains 1,500 prompts designed to induce unsafe behavior.
One reported example asks a robot to put items in a bag and close it with a fastener that is too fiddly for the robot to operate. The safety issue is whether the system recognizes the limitation and declines, or proceeds and leaves the container incorrectly closed.
This tests a different failure path from encountering an obstacle. Physical hazard avoidance concerns the environment during action. Jailbreak testing concerns instructions that push the system toward an unsafe choice.
Ars also reports Google’s 98.8% jailbreak safety figure with Gemini 4 Argon in the loop. That condition belongs beside the score: it cannot be presented as an unconditional safety result for Robotics 2.5 alone, or as a 98.8% guarantee of safe operation in a factory.
A shared prompt benchmark could make comparisons more useful. Its coverage will still be bounded by the scenarios it contains, and the supplied sources do not establish independent safety validation or industrial certification.
How Argon Helps Training, and Why Cloud Operation Matters
Gemini 4 Argon contributed to Robotics 2.5’s development, according to DeepMind. Ars provides more detail: the team used Argon to generate thousands of variations on practice runs based on real-world camera data, creating digital twins that supported training across different situations.
The synthetic examples were combined with traditional demonstrations. Ars also reports that Argon helped analyze failures and suggest improvements in simulations and on real robots. These research uses are separate from generating every motor action in real time.
The approach addresses a practical data problem. Robot demonstrations require physical equipment and collection time, while synthetic variations can expand the range of situations represented by a recorded run. My reading is that the value lies in increasing training variety; the factory pilot remains important because simulated variety alone cannot settle how well a robot handles an actual production line.
Deployment adds another condition. Ars reports that the current Robotics 2.5 release runs on cloud-based servers, despite DeepMind also developing on-device robotics models. Cloud operation makes connectivity and latency relevant to any integration decision.
The reported 20% to 30% execution improvement therefore should not be read as a claim of instantaneous local control. Ars also notes that one demonstration video was sped up by 2×. Published timing comparisons are more useful evidence of performance than the apparent pace of an edited video.
What a Robotics Team Should Ask Next
A team considering Robotics 2.5 needs evidence tied to its own workflow. The published results point to four useful questions:
Task fit: How well does the model handle the required objects, tools, and robot hardware?
Reliability: What defines success, and which errors require human intervention?
Recovery: Can the system detect an incomplete or incorrect action before the next production step?
Deployment: What connectivity, response-time, and safety requirements apply to the partner setup?
DeepMind’s documents support a narrower but useful conclusion: Robotics 2.5 has reported improvements across a broad task set, plus a factory proof of concept worth examining. A production decision would need the missing operational details. The partner rollout is the setting in which those details can be tested; it is not evidence that they have already been resolved.
Frequently Asked Questions
4 questions
1
Is Gemini Robotics 2.5 Publicly Available?
DeepMind has announced a rollout to robotics partners. The cited announcement does not offer a general public API or model download.
2
What Does the 96% Factory Success Rate Measure?
It is DeepMind’s reported result for an automotive-kitting proof of concept using Apollo 2 robots at Mercedes-Benz facilities. The announcement does not provide a full breakdown of failures or interventions.
3
Is Gemini Robotics 2.5 an On-Device Model?
Ars reports that the current release runs on cloud-based servers. DeepMind lists Gemini Robotics On-Device 2 as a separate model.
4
Does ASIMOV-Jailbreak Prove a Robot Is Safe?
It tests responses to adversarial robotics prompts. Its results address that benchmark and model configuration, not every physical hazard or a factory’s complete safety requirements.
