Gemini Robotics 2.5 Factory Performance Explained

Gemini Robotics 2.5 is a VLA model that connects vision and language to robot actions. The 96% success rate in the factory demonstration applies only to parts kitting. Decisions about actual deployment require further validation of failure recovery and operating conditions.

The performance figures in this article are based on Google's November 10, 2026 announcement.

How VLA and ER Work

VLA is a model that connects visual information and language instructions to robot actions. The name stands for Vision-Language-Action. ER stands for embodied reasoning, which plans tasks in physical environments.

The two models divide responsibilities in the factory demonstration. Gemini Robotics ER tracks the overall workflow. Gemini Robotics 2.5 controls both hands and arms.

The ER model is designed to determine what to do next and keep execution moving. It also determines whether a task is complete and whether a retry is needed. The official model information specifies that ER 2 outputs text. Google DeepMind's ER 2 description

The action model and the physical robot are not the same product. The official VLA description outlines a control scope ranging from arms to a humanoid's entire body. Performance on a specific robot must be interpreted alongside that hardware's conditions. Google DeepMind's VLA description

×

Task success rates and safety evaluations must be considered separately.

Task Evaluation Figures and Comparison Baselines

The announced improvement rates are results compared with the previous baseline model. They are not absolute success rates that apply to all tasks. The announcement explains that the extent of improvement varied by task type.

The following table separates the figures announced on November 10, 2026.

The task count refers to the types of tasks evaluated. This is different from the number of times each task was repeated. Statistical confidence therefore cannot be calculated from the task count alone. Google's Korean-language announcement

Examples of Relative Improvement Calculations

Converting success rate improvements into multiples makes their meaning clearer. The announced 30% improvement corresponds to 1.3 times the previous success rate. A 100% improvement means twice the previous success rate. These relationships are calculated from the announced figures.

Without the baseline success rate, the new absolute success rate cannot be calculated. A 100% improvement must not be read as success on every attempt. Speed improvements also cannot be converted into increases in total factory output.

The Scope of the 96% Factory Demonstration Success Rate

The announced 96% is the success rate for an automotive parts kitting demonstration. Kitting is the task of preparing parts for assembly in the required sequence. Apptronik's Apollo 2 performed this task in the demonstration.

The demonstration took place on actual Mercedes-Benz production lines. The announcement specifies Kecskemét, Hungary, and Sindelfingen, Germany. The robot picked metal brackets from supply bins and placed them in containers arranged by sequence.

A proof of concept, or PoC, is a test that checks feasibility in a specific environment. The success rate measured here applies within the scope of that test. It cannot be extended to the factory's overall automation rate or the success rate of every process.

Google outlined a use case in which robots handle repetitive preparation tasks. It explained that this helps technicians focus on complex assembly. It did not provide figures for workforce reductions or increases in labor productivity. Google DeepMind's factory demonstration description

Failure Recovery and Throughput Missing from the Success Rate

Decisions about factory deployment require the denominator of the success rate and the recovery time. Even with the same success rate, output can vary depending on how retries are handled. This analysis connects the announced figures to operational metrics.

These items are not additional verified achievements. They are measurement conditions that must be checked separately in on-site evaluations. The Korean-language announcement alone does not establish all the detailed conditions.

Good-unit throughput can also be calculated as an operational metric. The formula is “completed units that pass quality criteria ÷ total operating time.” Total operating time must include retries and recovery time. This is a proposed measurement, not a figure announced for the factory demonstration.

Robot operations also require management functions beyond the model. Apptronik describes Fleet Connect for status monitoring and task assignment. It shows that the scope of operations must also cover batteries and stopped states. Apptronik's Apollo 2 platform description

The Difference Between Physical Safety and Jailbreak Defense

Physical safety guidance and jailbreak defense evaluate different things. The former guides behavior to avoid nearby hazards during execution. The latter examines resistance to attacks that try to make the model follow dangerous instructions.

Inference is when a model receives input and makes a judgment. Google explains that it applies guidance at this stage to avoid hazards. Safety figures must also be distinguished by the model and evaluation they concern.

The following are the safety figures from the November 10, 2026 announcement.

The jailbreak defense rate is an evaluation result for Gemini 4 Argon. Citing it as the action model's accident-free rate in factories changes the subject of the evaluation. Improvement rates for safe task completion must also be distinguished from absolute success rates. Google DeepMind's safety evaluation description

The ASIMOV-Jailbreak repository distinguishes harmful instructions from legitimate instructions. It also includes evaluations of attacks that continue over multiple turns. The rate of excessive refusals of legitimate requests is another evaluation item. Google's ASIMOV evaluation repository

Physical devices also have separate safety features. Apptronik explains that the robot stops immediately when it detects an object in a configured collision zone. It also provides behavior settings based on surrounding zones. These are hardware features that must be checked separately from model evaluations.

Simulation and Real-World Training Data

According to Google, simulation was combined with real-world demonstration data. Gemini 4 Argon generates virtual training environments from camera scenes. It also supports research that develops hypotheses for improvement from videos of failures.

A digital twin recreates a real environment in virtual space. Virtual environments allow repeated tests under varied conditions. Tests with physical robots check whether those results also hold in physical environments. Google DeepMind's training process description

In this setup, factories serve as both use cases and evaluation sites. This is because real parts and task sequences test the model's scope of application. However, improvements in training efficiency cannot be directly converted into deployment cost savings.

The Difference Between Partner Availability and a Public API

The availability of the action model and the ER model must be checked separately. According to the announcement, the action model is being rolled out to robotics partners. The official ER 2 page specifies a public preview.

Google DeepMind's announcement described availability as follows.

Gemini Robotics 2.5 is rolling out to our robotics partners.

The source is the Gemini Robotics 2.5 announcement.

API access to ER does not mean access to the action model. The schedule for commercial deployment in South Korea also cannot be determined from this announcement alone. Each model's availability must be checked on its official model page.

Common Mistakes When Reading Performance Figures

Omitting the subject and unit of an evaluation mixes different achievements together. In particular, grouping figures by model name alone makes interpretation less clear. When citing figures, include the task and comparison conditions.

Key questions in this article

Separating model roles, evaluation units, and availability clarifies the key questions.

Is Gemini Robotics 2.5 a Physical Robot?

It is a VLA model that controls robots. Apollo 2 is a physical robot that uses it. Model performance must be distinguished from the robot's physical performance.

Are the Roughly 200 Tasks the Number of Processes in an Actual Factory?

The figure refers to the types of tasks in the announced model evaluation. It does not mean they are all automotive factory processes. The number of evaluation tasks and the scope of the factory demonstration are separate.

Can Deployment Benefits Be Calculated from the Factory Success Rate Alone?

Costs or output cannot be calculated from the success rate alone. Task time and failure recovery time must also be known. Quality assessment criteria and conditions for human intervention are also needed.

Does a Public Preview Mean the Same Thing as Commercial Deployment?

A public preview is an availability status that allows access to a specific model. It does not mean integration with factory systems is complete. The action model's availability conditions must also be checked separately.