Google DeepMind is Alphabet's London-headquartered AI laboratory. This profile concerns one part of it: the robotics group led by Carolina Parada, which since March 2025 has been shipping a family of "vision-language-action" foundation models under the Gemini Robotics name. The premise is that the control software for useful robots should not be written task by task, but learned — a general model that takes in camera images and plain-language instructions and outputs motor commands, portable across different robot bodies.
Parada presented Gemini Robotics on stage at the MACHINA Summit at Station F in Paris on 7 July 2026, the robotics event from which this tracker entry is drawn. The group has also been hiring from the hardware side: Aaron Saunders, formerly chief technology officer of Boston Dynamics, joined as vice-president of robotics hardware engineering in November 2025.
The release cadence has been rapid. The original Gemini Robotics model arrived in March 2025. In June 2025 came Gemini Robotics On-Device, a smaller variant that runs locally on the robot with no network connection and can be adapted to a new platform with 50 to 100 demonstrations; DeepMind showed it working on ALOHA research arms, the Franka FR3 and Apptronik's Apollo humanoid. September 2025 brought Gemini Robotics 1.5, which generates an internal reasoning sequence before acting and can transfer motions learned on one robot to a differently shaped one, alongside an "embodied reasoning" model, ER 1.5, available to any developer through the Gemini API.
On 30 July 2026, three weeks after MACHINA, the group announced Gemini Robotics 2, extending control from arms and hands to whole-body movement, including locomotion, with an On-Device 2 variant that adapts to new dual-arm robots with fewer than 200 real-world examples. The published results are honest about variance: one five-fingered hand achieved 92 per cent success removing a light bulb but only 36 per cent installing one. The full action models remain restricted to early-access partners, chiefly robot manufacturers; the reasoning models are publicly accessible. Since January 2026 Boston Dynamics and DeepMind have had a formal partnership to put Gemini Robotics models on the new electric Atlas humanoid.
The infrastructure case is indirect but strengthening. UK asset owners are already operating robots on critical sites: at Sellafield, Boston Dynamics Spot quadrupeds have moved from trials into routine radiological survey work, and the site completed its first teleoperated swabbing trial in February 2026. Today those machines run on task-specific autonomy or a human operator. Gemini Robotics is the most credible candidate for the general software layer that would let such fleets take on varied inspection and maintenance tasks without bespoke programming — and it is developed in London.
Two features matter specifically for infrastructure operators. Cross-embodiment transfer means capability developed on one platform is not locked to one vendor's hardware. And the on-device variants matter for nuclear, energy and water sites where connectivity is restricted or prohibited for security reasons. The Boston Dynamics partnership connects the models to hardware that is already inside the UK nuclear estate. It should be said plainly: there is no verified deployment of Gemini Robotics on UK or Irish infrastructure today.
The thesis is well evidenced on capability and proximity; what remains unproven is the landing. The confirming signals would be a Gemini Robotics early-access partner with a UK industrial channel, a trial on a UK asset, nuclear decommissioning being the obvious first candidate given the existing robot fleets, or regulators engaging with learned-control safety cases, which conventional nuclear and energy assurance regimes are not yet built for.
The gap between laboratory success rates and the near-total reliability infrastructure work demands is the central open question; 36 per cent success on a fiddly manipulation task is a research milestone, not an operational tool. Also worth watching is whether access stays confined to humanoid manufacturers or widens to the quadrupeds and arms that dominate real inspection work. The direction is clear; the UK connection beyond the postcode is what needs to materialise.
Sources: maddyness.com · deepmind.google · deepmind.google · gigazine.net · bostondynamics.com · therobotreport.com · theregister.com
Entries are verified against primary sources. The bar is the frontier: first-of-kind, new technology, new materials — delivery milestones of conventional infrastructure, however large, do not qualify. The frontier is tracked globally; every entry states what it means for UK and Irish infrastructure.
Verified by GridteamAI.
