Alphabet Inc.’s artificial intelligence research lab, Google DeepMind, Tuesday debuted the next generation of Gemini Robotics, the company’s most capable vision-language-action model to date for creating an “AI brain” for robotics.
Separately, DeepMind announced AlphaGenome 2 today, a next-generation AI model that can predict how DNA sequences and their variations affect human cell types, creating a powerful biological tool for fighting disease and drug development.
DeepMind said Gemini Robotics 2.5 enhances dexterity, reliability and safety to help engineers bring robots out of the lab and into dynamic, unstructured, real-world environments where they can coordinate alongside humans.
The company said that with the release of Gemini 4 Argon, the research division has been able to accelerate robot research and reach into physical AI. The new robotics AI model was tested across over 200 distinct tasks and the company said it anticipates scaling even further.
Argon, the company’s frontier artificial intelligence large language model, rolled out in late September, bringing advanced reasoning to the enterprise. Using the capabilities behind this new, flagship model, DeepMind said it allowed Gemini Robotics 2.5 to increase the breadth and speed with which robots can learn new skills and apply them to real-world scenarios.
The 2.5 version makes a significant jump in optimization and performance across three key areas. Boosted dexterity means robots running on 2.5 can perform more tasks with greater accuracy, including using everyday tools to handle delicate, flexible objects. With increased reliability, the new model successfully packed automotive parts with a 96% success rate during real-world proof-of-concept tests at Mercedes-Benz. Finally, DeepMind enhanced safety with improved physical steerability and introduced a new benchmark for testing against jailbreaks.
In evaluations, Robotics 2.5 now achieves a 30% to 100% success rate depending on the task and completes tasks 20% to 30% faster compared with the previous version’s baseline.
The company said that, like previous models, it natively works with many different types of robots, including Apptronik Inc.’s Apollo 2 humanoid robots and the Franka Duo platform with two-fingered parallel grippers.
“This is what we built Apollo for: useful work alongside people in real industrial settings,” Apptronik co-founder and Chief Executive Jeff Cardenas said. “Pairing Apollo 2 with Gemini Robotics, we saw the models improve dramatically in just a few weeks of post-training and carry over to Mercedes-Benz production with very little drop-off.”
Gemini Robotics 2.5 is rolling out today via DeepMind’s robotics partners, including Agile Robots SE, Apptronik and Boston Dynamics Inc.
AlphaGenome 2, bridging single-cell biology and human genetics
AI models such as AlphaGenome help drug development and biotechnology companies turn genetic data into testable leads. By predicting how DNA changes affect gene activity and which cells in the body those changes act in, they can guide researchers to promising drug targets and help prioritize experiments.
Announced today, AlphaGenome 2 adds finer detail about specific cell types and states, potentially revealing disease-related effects that broader tissue-level models miss. These predictions guide research rather than replace experimental validation; the model is not approved for clinical or diagnostic use.
The problem researchers face is that every cell in the body carries the same DNA; however, different genes are activated at different times to differentiate cells and adjust their behavior. For example, a heart muscle cell or an immune cell fighting an infection may have the same DNA, but they use the blueprint to produce different proteins, enzymes and other tools to do their job.
DeepMind used global efforts from the Human Cell Atlas, a scientific consortium that mapped gene expression in millions of cells, to give the world an unprecedented view of how different cells use genes. However, predicting the paths and variations that drive activation and expression changes remains a major challenge.
Professor Sarah Teichmann, co-founder of the Human Cell Atlas and professor of Stem Cell Medicine at the University of Cambridge, said what makes Alphagenome 2 so powerful is “how it adapts directly to new single-cell data without retraining, predicting how DNA variants act in newly profiled cell states,” and that it can adapt even from very small samples.
According to DeepMind, the new model can predict core experimental readouts of DNA function as the previous generation, including gene expression, histone modification and chromatin accessibility – but also adds DNA methylation, which is a significant part of epigenetics, which explains how behaviors, environment and life experiences can influence gene activity.
Image: Microsoft Designer/SiliconANGLE