Google DeepMind launches new robotics and gene prediction models

Moving robots into the real world

Google DeepMind has introduced Gemini Robotics 2.5, which serves as an advanced vision-language-action model. This technology acts as an "AI brain" for robots, allowing them to interpret visual information and translate it into physical actions. The goal is to move robots out of controlled laboratory settings and into dynamic, unstructured environments where they must work alongside humans.

The new model focuses on three primary improvements: dexterity, reliability, and safety. By improving dexterity, robots can now handle delicate or flexible objects and use everyday tools with higher accuracy. Reliability was tested in a real-world proof-of-concept with Mercedes-Benz, where the system successfully packed automotive parts with a 96% success rate.

Safety was also prioritized through improved physical steerability and a new benchmark to test against jailbreaks. According to SiliconANGLE, the model was tested across more than 200 distinct tasks. It is designed to work natively with various platforms, including the Apollo 2 humanoid robots from Apptronik and the Franka Duo platform.

Accelerating physical AI research

DeepMind utilized its frontier large language model, Gemini 4 Argon, to power the development of the new robotics system. The research division used the reasoning capabilities of Argon to help robots learn new skills more quickly. This approach allows for a broader range of applications and faster adaptation to different real-world scenarios.

The company noted that the 2.5 version represents a significant jump in optimization over previous iterations. In various evaluations, the new model achieved success rates between 30% and 100% depending on the specific task. It also completed those tasks 20% to 30% faster than the previous baseline established by the older version.

Partners like Boston Dynamics, Agile Robots SE, and Apptronik are already beginning to roll out the model. Apptronik co-founder Jeff Cardenas stated that the models improved dramatically in just a few weeks of post-training. This rapid adaptation allowed the technology to be carried over to production environments with very little drop-off in performance.

Predicting human cell behavior

In a separate announcement, DeepMind released AlphaGenome 2, a model designed to bridge single-cell biology and human genetics. Every cell in the human body contains the same DNA, but different genes are activated at different times to create different cell types. AlphaGenome 2 helps predict how specific DNA sequences and their variations affect these activations across different cell types.

This model is particularly useful for biotechnology and drug development companies. By predicting how DNA changes affect gene activity, researchers can identify promising drug targets more quickly. The model helps prioritize experiments by showing which cells in the body a specific genetic change might act in, potentially revealing disease-related effects that broader tissue-level models might miss.

To build this tool, DeepMind used data from the Human Cell Atlas, a scientific consortium that mapped gene expression in millions of cells. This provided an unprecedented view of how different cells use genes. While the model guides research, DeepMind clarified that it is not approved for clinical or diagnostic use and does not replace experimental validation.

Advancements in epigenetics

AlphaGenome 2 provides finer detail about specific cell types and states compared to its predecessor. It can predict core experimental readouts like gene expression and histone modification. Crucially, it also adds DNA methylation, which is a significant component of epigenetics. Epigenetics explains how environmental factors, life experiences, and behaviors can influence gene activity without changing the underlying DNA sequence.

A key strength of the model is its ability to adapt to new data without requiring full retraining. Professor Sarah Teichmann, a co-founder of the Human Cell Atlas, noted that the model can adapt even from very small samples. It can predict how DNA variants act in newly profiled cell states by adapting directly to new single-cell data.

By providing these high-resolution predictions, AlphaGenome 2 helps scientists understand the paths that drive expression changes. This allows for a more nuanced view of biology where the same blueprint is used to produce different proteins and enzymes. The model serves as a powerful tool for turning complex genetic data into testable leads for fighting disease.