{"content_id":"wnh4uzgh8d","slug":"gemini-robotics-2-5-vla-factory-performance","locale":"en","schema_type":"TechArticle","category":"ai_data","category_name":"AI Data","title":"Gemini Robotics 2.5 Factory Performance Explained","summary":"Gemini Robotics 2.5 is a model that connects vision and language to robot actions. This overview separately explains the scope of the 96% success rate in factory trials, performance improvement rates, safety evaluations, and each model's release status.","sponsorship_disclosure":null,"affiliate_disclosure":null,"commerce_disclosure":null,"author":{"name":"Injoys Editorial Team","url":"https://injoys.com/ko/about"},"key_points":["Gemini Robotics ER is a reasoning model that tracks workflows.","Gemini Robotics 2.5 controls the movements of both hands and arms in factory trials.","The 30~100% improvement in success rate is relative to the existing baseline model.","The 96% success rate in factory trials comes from automotive parts kitting tasks.","Partner access to the action model must be distinguished from the ER model's public preview."],"content_markdown":"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.\n\nThe performance figures in this article are based on Google's November 10, 2026 announcement.\n\n## How VLA and ER Work\n\nVLA 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.\n\nThe two models divide responsibilities in the factory demonstration. Gemini Robotics ER tracks the overall workflow. Gemini Robotics 2.5 controls both hands and arms.\n\n| Component | Role | Key distinction |\n| --- | --- | --- |\n| Visual information | Identifies objects and the surrounding scene | Input used to determine actions |\n| Language instructions | Communicate the task and its conditions | Input that expresses task goals |\n| ER model | Manages multistep plans and progress | High-level reasoning |\n| VLA action model | Generates robot movements | Control for physical actions |\n| Robot hardware | Executes movements and object manipulation | Physical devices such as hands, arms, and actuators |\n\nThe 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](https://deepmind.google/models/gemini-robotics/embodied-reasoning/)\n\nThe 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](https://deepmind.google/models/gemini-robotics/vla/)\n\n## Task Evaluation Figures and Comparison Baselines\n\nThe 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.\n\nThe following table separates the figures announced on November 10, 2026.\n\n| Item | Announcement | Scope of the figure |\n| --- | --- | --- |\n| Number of evaluation tasks | The introduction says more than 200; the detailed description says 200 | Understand the overall scale as around 200 |\n| Task success rate | 30~100% improvement over the previous baseline model | Relative improvement by task type |\n| Task completion speed | 20~30% improvement over the previous baseline model | Comparison of completion speeds for evaluation tasks |\n| Factory demonstration success rate | 96% | Apollo 2's automotive parts kitting demonstration |\n\nThe 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](https://blog.google/intl/ko-kr/company-news/technology/gemini-robotics-2-5-kr/)\n\n### Examples of Relative Improvement Calculations\n\nConverting 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.\n\n| Expression | Calculation | Information needed for interpretation |\n| --- | --- | --- |\n| 30% improvement in success rate | New success rate = previous success rate × 1.3 | Previous success rate for the same task |\n| 100% improvement in success rate | New success rate = previous success rate × 2 | Previous success rate for the same task |\n| Difference in success rates | New success rate - previous success rate | Both absolute success rates, measured in percentage points |\n\nWithout 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.\n\n## The Scope of the 96% Factory Demonstration Success Rate\n\nThe 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.\n\nThe 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.\n\nA 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.\n\nGoogle 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](https://deepmind.google/blog/gemini-robotics-2-5-accelerates-real-world-physical-intelligence/)\n\n## Failure Recovery and Throughput Missing from the Success Rate\n\nDecisions 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.\n\n| Item to check | What to verify | Effect on the decision |\n| --- | --- | --- |\n| Unit of success | Whether it is one part or an entire tray | The scope of work represented by the success rate |\n| Number of attempts | Total evaluation attempts and their distribution by task | Variability and representativeness of the results |\n| Retry handling | Whether success means first-attempt success or includes retries | Level of autonomous performance |\n| Human intervention | Number and timing of interventions | Required operating staff |\n| Failure recovery | Recovery time and conditions for resuming work | Throughput lost to interruptions |\n| Quality assessment | Criteria for positional errors and part damage | Whether the output can be used in subsequent assembly |\n| Continuous operation | Operating time and causes of stops | Stability in actual operation |\n\nThese 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.\n\nGood-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.\n\nRobot 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](https://apptronik.com/apollo/apollo-2)\n\n## The Difference Between Physical Safety and Jailbreak Defense\n\nPhysical 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.\n\nInference 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.\n\nThe following are the safety figures from the November 10, 2026 announcement.\n\n| Evaluation | Announced result | Scope of interpretation |\n| --- | --- | --- |\n| Safe task completion around obstacles | 33% improvement | Improvement from applying guidance |\n| Safe task completion with people nearby | 43% improvement | Improvement from applying guidance |\n| Gemini 4 Argon jailbreak defense rate | 99.5% | Announced jailbreak evaluation result |\n| Baseline model's jailbreak defense rate | 82% | Comparison result for Gemini Robotics ER 2 |\n\nThe 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](https://deepmind.google/blog/gemini-robotics-2-5-accelerates-real-world-physical-intelligence/)\n\nThe 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](https://huggingface.co/datasets/google/asimov_agentic)\n\nPhysical 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.\n\n## Simulation and Real-World Training Data\n\nAccording 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.\n\nA 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](https://deepmind.google/blog/gemini-robotics-2-5-accelerates-real-world-physical-intelligence/)\n\nIn 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.\n\n## The Difference Between Partner Availability and a Public API\n\nThe 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.\n\nGoogle DeepMind's announcement described availability as follows.\n\n\u003e Gemini Robotics 2.5 is rolling out to our robotics partners.\n\nThe source is the [Gemini Robotics 2.5 announcement](https://deepmind.google/blog/gemini-robotics-2-5-accelerates-real-world-physical-intelligence/).\n\n| Model | Confirmed availability | Key distinction |\n| --- | --- | --- |\n| Gemini Robotics 2.5 | Rolling out to robotics partners | Distinct from releasing an action API for general developers |\n| Gemini Robotics ER 2 | Public preview in Google AI Studio and the Gemini API | Scope of access to the reasoning model |\n\nAPI 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.\n\n## Common Mistakes When Reading Performance Figures\n\nOmitting 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.\n\n| Common interpretation | Correct distinction |\n| --- | --- |\n| A 100% improvement in success rate means every task succeeded | Relative improvement and absolute success rate are different |\n| The factory success rate is the factory's overall automation rate | Kitting PoC results and the scope of automation are different |\n| The jailbreak defense rate is the robot's accident-free rate | Attack resistance and physical accident metrics are different |\n| The ER API provides immediate access to the action model too | Access permissions must be checked separately for each model |\n| Output increases by the same rate as task completion speed | Measurements must include recovery time and process waiting time |\n\n## Frequently Asked Questions\n\nSeparating model roles, evaluation units, and availability clarifies the key questions.\n\n### Is Gemini Robotics 2.5 a Physical Robot?\n\nIt 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.\n\n### Are the Roughly 200 Tasks the Number of Processes in an Actual Factory?\n\nThe 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.\n\n### Can Deployment Benefits Be Calculated from the Factory Success Rate Alone?\n\nCosts 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.\n\n### Does a Public Preview Mean the Same Thing as Commercial Deployment?\n\nA 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.","content_html":"\u003cp\u003eGemini 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.\u003c/p\u003e\n\u003cp\u003eThe performance figures in this article are based on Google's November 10, 2026 announcement.\u003c/p\u003e\n\u003ch2\u003e\n\u003ca href=\"#how-vla-and-er-work\" class=\"anchor\" id=\"how-vla-and-er-work\"\u003e\u003c/a\u003eHow VLA and ER Work\u003c/h2\u003e\n\u003cp\u003eVLA 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.\u003c/p\u003e\n\u003cp\u003eThe two models divide responsibilities in the factory demonstration. Gemini Robotics ER tracks the overall workflow. Gemini Robotics 2.5 controls both hands and arms.\u003c/p\u003e\n\u003cdiv class=\"overflow-x-auto\"\u003e\u003ctable\u003e\n\u003cthead\u003e\n\u003ctr\u003e\n\u003cth\u003eComponent\u003c/th\u003e\n\u003cth\u003eRole\u003c/th\u003e\n\u003cth\u003eKey distinction\u003c/th\u003e\n\u003c/tr\u003e\n\u003c/thead\u003e\n\u003ctbody\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Component\"\u003eVisual information\u003c/td\u003e\n\u003ctd data-label=\"Role\"\u003eIdentifies objects and the surrounding scene\u003c/td\u003e\n\u003ctd data-label=\"Key distinction\"\u003eInput used to determine actions\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Component\"\u003eLanguage instructions\u003c/td\u003e\n\u003ctd data-label=\"Role\"\u003eCommunicate the task and its conditions\u003c/td\u003e\n\u003ctd data-label=\"Key distinction\"\u003eInput that expresses task goals\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Component\"\u003eER model\u003c/td\u003e\n\u003ctd data-label=\"Role\"\u003eManages multistep plans and progress\u003c/td\u003e\n\u003ctd data-label=\"Key distinction\"\u003eHigh-level reasoning\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Component\"\u003eVLA action model\u003c/td\u003e\n\u003ctd data-label=\"Role\"\u003eGenerates robot movements\u003c/td\u003e\n\u003ctd data-label=\"Key distinction\"\u003eControl for physical actions\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Component\"\u003eRobot hardware\u003c/td\u003e\n\u003ctd data-label=\"Role\"\u003eExecutes movements and object manipulation\u003c/td\u003e\n\u003ctd data-label=\"Key distinction\"\u003ePhysical devices such as hands, arms, and actuators\u003c/td\u003e\n\u003c/tr\u003e\n\u003c/tbody\u003e\n\u003c/table\u003e\u003c/div\u003e\n\u003cp\u003eThe 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. \u003ca href=\"https://deepmind.google/models/gemini-robotics/embodied-reasoning/\"\u003eGoogle DeepMind's ER 2 description\u003c/a\u003e\u003c/p\u003e\n\u003cp\u003eThe 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. \u003ca href=\"https://deepmind.google/models/gemini-robotics/vla/\"\u003eGoogle DeepMind's VLA description\u003c/a\u003e\u003c/p\u003e\n\u003ch2\u003e\n\u003ca href=\"#task-evaluation-figures-and-comparison-baselines\" class=\"anchor\" id=\"task-evaluation-figures-and-comparison-baselines\"\u003e\u003c/a\u003eTask Evaluation Figures and Comparison Baselines\u003c/h2\u003e\n\u003cp\u003eThe 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.\u003c/p\u003e\n\u003cp\u003eThe following table separates the figures announced on November 10, 2026.\u003c/p\u003e\n\u003cdiv class=\"overflow-x-auto\"\u003e\u003ctable\u003e\n\u003cthead\u003e\n\u003ctr\u003e\n\u003cth\u003eItem\u003c/th\u003e\n\u003cth\u003eAnnouncement\u003c/th\u003e\n\u003cth\u003eScope of the figure\u003c/th\u003e\n\u003c/tr\u003e\n\u003c/thead\u003e\n\u003ctbody\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Item\"\u003eNumber of evaluation tasks\u003c/td\u003e\n\u003ctd data-label=\"Announcement\"\u003eThe introduction says more than 200; the detailed description says 200\u003c/td\u003e\n\u003ctd data-label=\"Scope of the figure\"\u003eUnderstand the overall scale as around 200\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Item\"\u003eTask success rate\u003c/td\u003e\n\u003ctd data-label=\"Announcement\"\u003e30~100% improvement over the previous baseline model\u003c/td\u003e\n\u003ctd data-label=\"Scope of the figure\"\u003eRelative improvement by task type\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Item\"\u003eTask completion speed\u003c/td\u003e\n\u003ctd data-label=\"Announcement\"\u003e20~30% improvement over the previous baseline model\u003c/td\u003e\n\u003ctd data-label=\"Scope of the figure\"\u003eComparison of completion speeds for evaluation tasks\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Item\"\u003eFactory demonstration success rate\u003c/td\u003e\n\u003ctd data-label=\"Announcement\"\u003e96%\u003c/td\u003e\n\u003ctd data-label=\"Scope of the figure\"\u003eApollo 2's automotive parts kitting demonstration\u003c/td\u003e\n\u003c/tr\u003e\n\u003c/tbody\u003e\n\u003c/table\u003e\u003c/div\u003e\n\u003cp\u003eThe 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. \u003ca href=\"https://blog.google/intl/ko-kr/company-news/technology/gemini-robotics-2-5-kr/\"\u003eGoogle's Korean-language announcement\u003c/a\u003e\u003c/p\u003e\n\u003ch3\u003e\n\u003ca href=\"#examples-of-relative-improvement-calculations\" class=\"anchor\" id=\"examples-of-relative-improvement-calculations\"\u003e\u003c/a\u003eExamples of Relative Improvement Calculations\u003c/h3\u003e\n\u003cp\u003eConverting 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.\u003c/p\u003e\n\u003cdiv class=\"overflow-x-auto\"\u003e\u003ctable\u003e\n\u003cthead\u003e\n\u003ctr\u003e\n\u003cth\u003eExpression\u003c/th\u003e\n\u003cth\u003eCalculation\u003c/th\u003e\n\u003cth\u003eInformation needed for interpretation\u003c/th\u003e\n\u003c/tr\u003e\n\u003c/thead\u003e\n\u003ctbody\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Expression\"\u003e30% improvement in success rate\u003c/td\u003e\n\u003ctd data-label=\"Calculation\"\u003eNew success rate = previous success rate × 1.3\u003c/td\u003e\n\u003ctd data-label=\"Information needed for interpretation\"\u003ePrevious success rate for the same task\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Expression\"\u003e100% improvement in success rate\u003c/td\u003e\n\u003ctd data-label=\"Calculation\"\u003eNew success rate = previous success rate × 2\u003c/td\u003e\n\u003ctd data-label=\"Information needed for interpretation\"\u003ePrevious success rate for the same task\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Expression\"\u003eDifference in success rates\u003c/td\u003e\n\u003ctd data-label=\"Calculation\"\u003eNew success rate - previous success rate\u003c/td\u003e\n\u003ctd data-label=\"Information needed for interpretation\"\u003eBoth absolute success rates, measured in percentage points\u003c/td\u003e\n\u003c/tr\u003e\n\u003c/tbody\u003e\n\u003c/table\u003e\u003c/div\u003e\n\u003cp\u003eWithout 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.\u003c/p\u003e\n\u003ch2\u003e\n\u003ca href=\"#the-scope-of-the-96-factory-demonstration-success-rate\" class=\"anchor\" id=\"the-scope-of-the-96-factory-demonstration-success-rate\"\u003e\u003c/a\u003eThe Scope of the 96% Factory Demonstration Success Rate\u003c/h2\u003e\n\u003cp\u003eThe announced \u003cstrong\u003e96%\u003c/strong\u003e 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.\u003c/p\u003e\n\u003cp\u003eThe 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.\u003c/p\u003e\n\u003cp\u003eA 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.\u003c/p\u003e\n\u003cp\u003eGoogle 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. \u003ca href=\"https://deepmind.google/blog/gemini-robotics-2-5-accelerates-real-world-physical-intelligence/\"\u003eGoogle DeepMind's factory demonstration description\u003c/a\u003e\u003c/p\u003e\n\u003ch2\u003e\n\u003ca href=\"#failure-recovery-and-throughput-missing-from-the-success-rate\" class=\"anchor\" id=\"failure-recovery-and-throughput-missing-from-the-success-rate\"\u003e\u003c/a\u003eFailure Recovery and Throughput Missing from the Success Rate\u003c/h2\u003e\n\u003cp\u003eDecisions 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.\u003c/p\u003e\n\u003cdiv class=\"overflow-x-auto\"\u003e\u003ctable\u003e\n\u003cthead\u003e\n\u003ctr\u003e\n\u003cth\u003eItem to check\u003c/th\u003e\n\u003cth\u003eWhat to verify\u003c/th\u003e\n\u003cth\u003eEffect on the decision\u003c/th\u003e\n\u003c/tr\u003e\n\u003c/thead\u003e\n\u003ctbody\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Item to check\"\u003eUnit of success\u003c/td\u003e\n\u003ctd data-label=\"What to verify\"\u003eWhether it is one part or an entire tray\u003c/td\u003e\n\u003ctd data-label=\"Effect on the decision\"\u003eThe scope of work represented by the success rate\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Item to check\"\u003eNumber of attempts\u003c/td\u003e\n\u003ctd data-label=\"What to verify\"\u003eTotal evaluation attempts and their distribution by task\u003c/td\u003e\n\u003ctd data-label=\"Effect on the decision\"\u003eVariability and representativeness of the results\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Item to check\"\u003eRetry handling\u003c/td\u003e\n\u003ctd data-label=\"What to verify\"\u003eWhether success means first-attempt success or includes retries\u003c/td\u003e\n\u003ctd data-label=\"Effect on the decision\"\u003eLevel of autonomous performance\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Item to check\"\u003eHuman intervention\u003c/td\u003e\n\u003ctd data-label=\"What to verify\"\u003eNumber and timing of interventions\u003c/td\u003e\n\u003ctd data-label=\"Effect on the decision\"\u003eRequired operating staff\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Item to check\"\u003eFailure recovery\u003c/td\u003e\n\u003ctd data-label=\"What to verify\"\u003eRecovery time and conditions for resuming work\u003c/td\u003e\n\u003ctd data-label=\"Effect on the decision\"\u003eThroughput lost to interruptions\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Item to check\"\u003eQuality assessment\u003c/td\u003e\n\u003ctd data-label=\"What to verify\"\u003eCriteria for positional errors and part damage\u003c/td\u003e\n\u003ctd data-label=\"Effect on the decision\"\u003eWhether the output can be used in subsequent assembly\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Item to check\"\u003eContinuous operation\u003c/td\u003e\n\u003ctd data-label=\"What to verify\"\u003eOperating time and causes of stops\u003c/td\u003e\n\u003ctd data-label=\"Effect on the decision\"\u003eStability in actual operation\u003c/td\u003e\n\u003c/tr\u003e\n\u003c/tbody\u003e\n\u003c/table\u003e\u003c/div\u003e\n\u003cp\u003eThese 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.\u003c/p\u003e\n\u003cp\u003eGood-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.\u003c/p\u003e\n\u003cp\u003eRobot 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. \u003ca href=\"https://apptronik.com/apollo/apollo-2\"\u003eApptronik's Apollo 2 platform description\u003c/a\u003e\u003c/p\u003e\n\u003ch2\u003e\n\u003ca href=\"#the-difference-between-physical-safety-and-jailbreak-defense\" class=\"anchor\" id=\"the-difference-between-physical-safety-and-jailbreak-defense\"\u003e\u003c/a\u003eThe Difference Between Physical Safety and Jailbreak Defense\u003c/h2\u003e\n\u003cp\u003ePhysical 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.\u003c/p\u003e\n\u003cp\u003eInference 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.\u003c/p\u003e\n\u003cp\u003eThe following are the safety figures from the November 10, 2026 announcement.\u003c/p\u003e\n\u003cdiv class=\"overflow-x-auto\"\u003e\u003ctable\u003e\n\u003cthead\u003e\n\u003ctr\u003e\n\u003cth\u003eEvaluation\u003c/th\u003e\n\u003cth\u003eAnnounced result\u003c/th\u003e\n\u003cth\u003eScope of interpretation\u003c/th\u003e\n\u003c/tr\u003e\n\u003c/thead\u003e\n\u003ctbody\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Evaluation\"\u003eSafe task completion around obstacles\u003c/td\u003e\n\u003ctd data-label=\"Announced result\"\u003e33% improvement\u003c/td\u003e\n\u003ctd data-label=\"Scope of interpretation\"\u003eImprovement from applying guidance\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Evaluation\"\u003eSafe task completion with people nearby\u003c/td\u003e\n\u003ctd data-label=\"Announced result\"\u003e43% improvement\u003c/td\u003e\n\u003ctd data-label=\"Scope of interpretation\"\u003eImprovement from applying guidance\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Evaluation\"\u003eGemini 4 Argon jailbreak defense rate\u003c/td\u003e\n\u003ctd data-label=\"Announced result\"\u003e99.5%\u003c/td\u003e\n\u003ctd data-label=\"Scope of interpretation\"\u003eAnnounced jailbreak evaluation result\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Evaluation\"\u003eBaseline model's jailbreak defense rate\u003c/td\u003e\n\u003ctd data-label=\"Announced result\"\u003e82%\u003c/td\u003e\n\u003ctd data-label=\"Scope of interpretation\"\u003eComparison result for Gemini Robotics ER 2\u003c/td\u003e\n\u003c/tr\u003e\n\u003c/tbody\u003e\n\u003c/table\u003e\u003c/div\u003e\n\u003cp\u003eThe 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. \u003ca href=\"https://deepmind.google/blog/gemini-robotics-2-5-accelerates-real-world-physical-intelligence/\"\u003eGoogle DeepMind's safety evaluation description\u003c/a\u003e\u003c/p\u003e\n\u003cp\u003eThe 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. \u003ca href=\"https://huggingface.co/datasets/google/asimov_agentic\"\u003eGoogle's ASIMOV evaluation repository\u003c/a\u003e\u003c/p\u003e\n\u003cp\u003ePhysical 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.\u003c/p\u003e\n\u003ch2\u003e\n\u003ca href=\"#simulation-and-real-world-training-data\" class=\"anchor\" id=\"simulation-and-real-world-training-data\"\u003e\u003c/a\u003eSimulation and Real-World Training Data\u003c/h2\u003e\n\u003cp\u003eAccording 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.\u003c/p\u003e\n\u003cp\u003eA 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. \u003ca href=\"https://deepmind.google/blog/gemini-robotics-2-5-accelerates-real-world-physical-intelligence/\"\u003eGoogle DeepMind's training process description\u003c/a\u003e\u003c/p\u003e\n\u003cp\u003eIn 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.\u003c/p\u003e\n\u003ch2\u003e\n\u003ca href=\"#the-difference-between-partner-availability-and-a-public-api\" class=\"anchor\" id=\"the-difference-between-partner-availability-and-a-public-api\"\u003e\u003c/a\u003eThe Difference Between Partner Availability and a Public API\u003c/h2\u003e\n\u003cp\u003eThe 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.\u003c/p\u003e\n\u003cp\u003eGoogle DeepMind's announcement described availability as follows.\u003c/p\u003e\n\u003cblockquote\u003e\n\u003cp\u003eGemini Robotics 2.5 is rolling out to our robotics partners.\u003c/p\u003e\n\u003c/blockquote\u003e\n\u003cp\u003eThe source is the \u003ca href=\"https://deepmind.google/blog/gemini-robotics-2-5-accelerates-real-world-physical-intelligence/\"\u003eGemini Robotics 2.5 announcement\u003c/a\u003e.\u003c/p\u003e\n\u003cdiv class=\"overflow-x-auto\"\u003e\u003ctable\u003e\n\u003cthead\u003e\n\u003ctr\u003e\n\u003cth\u003eModel\u003c/th\u003e\n\u003cth\u003eConfirmed availability\u003c/th\u003e\n\u003cth\u003eKey distinction\u003c/th\u003e\n\u003c/tr\u003e\n\u003c/thead\u003e\n\u003ctbody\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Model\"\u003eGemini Robotics 2.5\u003c/td\u003e\n\u003ctd data-label=\"Confirmed availability\"\u003eRolling out to robotics partners\u003c/td\u003e\n\u003ctd data-label=\"Key distinction\"\u003eDistinct from releasing an action API for general developers\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Model\"\u003eGemini Robotics ER 2\u003c/td\u003e\n\u003ctd data-label=\"Confirmed availability\"\u003ePublic preview in Google AI Studio and the Gemini API\u003c/td\u003e\n\u003ctd data-label=\"Key distinction\"\u003eScope of access to the reasoning model\u003c/td\u003e\n\u003c/tr\u003e\n\u003c/tbody\u003e\n\u003c/table\u003e\u003c/div\u003e\n\u003cp\u003eAPI 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.\u003c/p\u003e\n\u003ch2\u003e\n\u003ca href=\"#common-mistakes-when-reading-performance-figures\" class=\"anchor\" id=\"common-mistakes-when-reading-performance-figures\"\u003e\u003c/a\u003eCommon Mistakes When Reading Performance Figures\u003c/h2\u003e\n\u003cp\u003eOmitting 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.\u003c/p\u003e\n\u003cdiv class=\"overflow-x-auto\"\u003e\u003ctable\u003e\n\u003cthead\u003e\n\u003ctr\u003e\n\u003cth\u003eCommon interpretation\u003c/th\u003e\n\u003cth\u003eCorrect distinction\u003c/th\u003e\n\u003c/tr\u003e\n\u003c/thead\u003e\n\u003ctbody\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Common interpretation\"\u003eA 100% improvement in success rate means every task succeeded\u003c/td\u003e\n\u003ctd data-label=\"Correct distinction\"\u003eRelative improvement and absolute success rate are different\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Common interpretation\"\u003eThe factory success rate is the factory's overall automation rate\u003c/td\u003e\n\u003ctd data-label=\"Correct distinction\"\u003eKitting PoC results and the scope of automation are different\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Common interpretation\"\u003eThe jailbreak defense rate is the robot's accident-free rate\u003c/td\u003e\n\u003ctd data-label=\"Correct distinction\"\u003eAttack resistance and physical accident metrics are different\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Common interpretation\"\u003eThe ER API provides immediate access to the action model too\u003c/td\u003e\n\u003ctd data-label=\"Correct distinction\"\u003eAccess permissions must be checked separately for each model\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Common interpretation\"\u003eOutput increases by the same rate as task completion speed\u003c/td\u003e\n\u003ctd data-label=\"Correct distinction\"\u003eMeasurements must include recovery time and process waiting time\u003c/td\u003e\n\u003c/tr\u003e\n\u003c/tbody\u003e\n\u003c/table\u003e\u003c/div\u003e\n\u003ch2\u003e\n\u003ca href=\"#frequently-asked-questions\" class=\"anchor\" id=\"frequently-asked-questions\"\u003e\u003c/a\u003eFrequently Asked Questions\u003c/h2\u003e\n\u003cp\u003eSeparating model roles, evaluation units, and availability clarifies the key questions.\u003c/p\u003e\n\u003ch3\u003e\n\u003ca href=\"#is-gemini-robotics-25-a-physical-robot\" class=\"anchor\" id=\"is-gemini-robotics-25-a-physical-robot\"\u003e\u003c/a\u003eIs Gemini Robotics 2.5 a Physical Robot?\u003c/h3\u003e\n\u003cp\u003eIt 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.\u003c/p\u003e\n\u003ch3\u003e\n\u003ca href=\"#are-the-roughly-200-tasks-the-number-of-processes-in-an-actual-factory\" class=\"anchor\" id=\"are-the-roughly-200-tasks-the-number-of-processes-in-an-actual-factory\"\u003e\u003c/a\u003eAre the Roughly 200 Tasks the Number of Processes in an Actual Factory?\u003c/h3\u003e\n\u003cp\u003eThe 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.\u003c/p\u003e\n\u003ch3\u003e\n\u003ca href=\"#can-deployment-benefits-be-calculated-from-the-factory-success-rate-alone\" class=\"anchor\" id=\"can-deployment-benefits-be-calculated-from-the-factory-success-rate-alone\"\u003e\u003c/a\u003eCan Deployment Benefits Be Calculated from the Factory Success Rate Alone?\u003c/h3\u003e\n\u003cp\u003eCosts 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.\u003c/p\u003e\n\u003ch3\u003e\n\u003ca href=\"#does-a-public-preview-mean-the-same-thing-as-commercial-deployment\" class=\"anchor\" id=\"does-a-public-preview-mean-the-same-thing-as-commercial-deployment\"\u003e\u003c/a\u003eDoes a Public Preview Mean the Same Thing as Commercial Deployment?\u003c/h3\u003e\n\u003cp\u003eA 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.\u003c/p\u003e\n","tags":["Generative AI","AI Development","Robots","Technology strategy"],"faqs":[{"question":"What is the VLA in Gemini Robotics 2.5?","answer":"VLA stands for Vision-Language-Action. It refers to a model that connects visual information and language instructions to robot actions."},{"question":"How does Gemini Robotics ER differ from the action model?","answer":"ER is a reasoning model that manages task planning and progress. In the factory demonstration, the action model controlled both hands and arms."},{"question":"Does the 30~100% improvement in success rate refer to an absolute success rate?","answer":"It is a relative improvement compared with the previous baseline model. Calculating the new absolute success rate requires the previous success rate for the same task."},{"question":"What does the 96% success rate in the factory demonstration measure?","answer":"It is the success rate reported for an automotive parts kitting PoC. It reflects the results of tasks performed by Apollo 2 on a Mercedes-Benz production line."},{"question":"Can faster task completion be taken to mean increased production output?","answer":"The overall rate of increase in production output cannot be calculated from completion speed alone. Operating time that includes retries, failure recovery, and process wait times is needed."},{"question":"Which model achieved the 99.5% jailbreak defense rate?","answer":"It is the result of the jailbreak evaluation for Gemini 4 Argon announced by Google. It does not refer to the action model's accident-free rate in factories."},{"question":"Can regular developers use the action model API right away?","answer":"The announced availability of the action model is limited to robotics partners. The public preview of ER 2 must be distinguished from access to the action model."},{"question":"Has a timeline been announced for commercial deployment in domestic factories?","answer":"A timeline for domestic commercial deployment cannot be confirmed based on this announcement alone. Specific availability conditions and timelines must be checked with Google and robot suppliers."}],"sources":[{"url":"https://blog.google/intl/ko-kr/company-news/technology/gemini-robotics-2-5-kr/","title":"Google's Korean announcement: Gemini Robotics 2.5: Accelerating physical intelligence in the real world","type":"source"},{"url":"https://deepmind.google/blog/gemini-robotics-2-5-accelerates-real-world-physical-intelligence/","title":"Google DeepMind: Announcing Gemini Robotics 2.5","type":"source"},{"url":"https://deepmind.google/models/gemini-robotics/vla/","title":"Google DeepMind: Gemini Robotics 2.5 VLA model","type":"source"},{"url":"https://deepmind.google/models/gemini-robotics/embodied-reasoning/","title":"Google DeepMind: Gemini Robotics ER 2 model information","type":"source"},{"url":"https://apptronik.com/apollo/apollo-2","title":"Apptronik: Apollo 2 platform and operational and safety features","type":"source"},{"url":"https://huggingface.co/datasets/google/asimov_agentic","title":"Google: Asimov Agentic Safety Evaluation and ASIMOV-Jailbreak","type":"source"}],"images":[{"id":2720,"url":"https://injoys.com/rails/active_storage/blobs/proxy/eyJfcmFpbHMiOnsiZGF0YSI6NDE3NzQsInB1ciI6ImJsb2JfaWQifX0=--942c42c606bce2666318996725dd445f966e877a/ai-4a3f957c.webp","is_representative":true,"generation_method":"ai_photo","license":"ai_generated","mime_type":"image/webp","width":1536,"height":1024,"translations":{"ko":{"alt":"작업자가 안전 유리 너머로 지켜보는 가운데 양팔 로봇이 금속 브래킷을 칸막이 트레이에 넣고 있다.","caption":"공장 실증 성공률 96%는 자동차 부품 키팅 작업에 한정된 결과다.","description":null},"en":{"alt":"A dual-arm robot places metal brackets in a divided tray as a worker watches from behind a safety barrier.","caption":"The 96% factory trial success rate applies specifically to automotive parts kitting.","description":null},"ja":{"alt":"作業員が安全ガラス越しに見守る中、双腕ロボットが金属ブラケットを仕切り付きトレーに入れている。","caption":"工場実証での成功率96％は、自動車部品のキッティング作業に限定された結果だ。","description":null},"es":{"alt":"Un robot de dos brazos coloca soportes metálicos en una bandeja mientras una operaria observa tras un cristal.","caption":"La tasa de éxito del 96% en la prueba de fábrica corresponde específicamente a la preparación de kits de piezas de automóvil.","description":null},"id":{"alt":"Robot dua lengan menaruh braket logam ke baki bersekat, sementara pekerja mengamati dari balik kaca pengaman.","caption":"Tingkat keberhasilan 96% dalam uji pabrik khusus berlaku untuk penyiapan kit komponen otomotif.","description":null},"pt":{"alt":"Um robô de dois braços coloca suportes metálicos numa bandeja enquanto uma operadora observa atrás de um vidro.","caption":"A taxa de sucesso de 96% no teste de fábrica se refere especificamente à preparação de kits de peças automotivas.","description":null},"zh-hant":{"alt":"一名作業員隔著安全玻璃觀看，雙臂機器人將金屬支架放入分隔托盤。","caption":"工廠實證的96%成功率僅適用於汽車零件配套作業。","description":null},"de":{"alt":"Ein zweiarmiger Roboter legt Metallwinkel in eine Fächerschale, während eine Fachkraft hinter Schutzglas zusieht.","caption":"Die Erfolgsquote von 96 % im Werkstest gilt speziell für die Zusammenstellung von Automobilteilesätzen.","description":null}}},{"id":2721,"url":"https://injoys.com/rails/active_storage/blobs/proxy/eyJfcmFpbHMiOnsiZGF0YSI6NDE3ODAsInB1ciI6ImJsb2JfaWQifX0=--19b385253034a3ff7240ac9fd1708a09552dc8d6/ai-abdf1234.webp","is_representative":false,"generation_method":"ai_semi","license":"ai_generated","mime_type":"image/webp","width":1536,"height":1024,"translations":{"ko":{"alt":"보호안경과 주황색 작업복을 착용한 엔지니어가 로봇 집게와 완충재 장애물 사이의 간격을 살핀다.","caption":"작업 성공률과 안전성 평가는 구분해서 살펴봐야 한다.","description":null},"en":{"alt":"An engineer in safety glasses and an orange jacket checks the gap between a robot gripper and a padded obstacle.","caption":"Task success rates and safety evaluations must be considered separately.","description":null},"ja":{"alt":"保護メガネとオレンジ色の作業着を着た技術者が、ロボットのグリッパーと緩衝材付き障害物の間隔を確認する。","caption":"作業成功率と安全性評価は分けて捉える必要がある。","description":null},"es":{"alt":"Una ingeniera con gafas protectoras revisa la separación entre una pinza robótica y un obstáculo acolchado.","caption":"La tasa de éxito de las tareas y la evaluación de seguridad deben considerarse por separado.","description":null},"id":{"alt":"Insinyur berkacamata pelindung memeriksa jarak antara pencengkeram robot dan penghalang berlapis busa.","caption":"Tingkat keberhasilan tugas dan evaluasi keselamatan perlu ditinjau secara terpisah.","description":null},"pt":{"alt":"Uma engenheira de óculos de proteção verifica a distância entre uma garra robótica e um obstáculo acolchoado.","caption":"A taxa de sucesso das tarefas e a avaliação de segurança devem ser consideradas separadamente.","description":null},"zh-hant":{"alt":"一名戴著護目鏡、身穿橘色工作服的工程師，檢查機器人夾爪與軟墊障礙物之間的間距。","caption":"任務成功率與安全性評估應分開看待。","description":null},"de":{"alt":"Eine Ingenieurin mit Schutzbrille prüft den Abstand zwischen Robotergreifer und gepolstertem Hindernis.","caption":"Aufgabenerfolgsquoten und Sicherheitsbewertungen müssen getrennt betrachtet werden.","description":null}}}],"published_at":"2026-11-11T22:26:13+09:00","updated_at":"2026-11-11T22:26:13+09:00","license":"cc_by","translation_status":"reviewed","available_locales":["ko","en","ja","es"],"data_locales":["ko","en","ja","es","id","pt","zh-hant","de"],"url":"https://injoys.com/en/articles/gemini-robotics-2-5-vla-factory-performance"}