Google DeepMind is the AI research lab behind Gemini. In June 2026 it launched its first European robotics accelerator, selecting 15 early-stage startups for a three-month, equity-free programme with access to Gemini Robotics models and up to $350,000 in Google Cloud credits. In July 2026, with Isomorphic Labs, it published a bioresilience strategy adapting its SynthID watermarking to screen AI-generated biological sequences at DNA synthesis facilities, alongside AlphaEvolve and AlphaGenome tools aimed at faster outbreak detection. At the end of that month it released Gemini Robotics ER 2, a high-level reasoning layer that tracks task progress from live video and coordinates several robots in one workspace, reaching European developers through the Gemini API rather than as open weights. In September 2026 a London DeepMind team published SynthIDBio in Nature, a method that watermarks AI-designed protein sequences and structures without reducing binding in lab tests, though resequencing with an open design tool removes the mark.
In October 2026 its AlphaProof Nexus system, which writes Lean proofs with Gemini 3.1 Pro, was published in Science after solving nine open Erdős problems and proving 44 OEIS conjectures.
Later in October a DeepMind team reported that four agents running on Gemini 4 Argon beat Minecraft without human help in 27 hours, coordinating through chat and a shared log of their mistakes.
On 30 October 2026 it published pilot results for PlanAssist, a Gemini-based planning tool that cut officer assessment time by 58% across three English councils.