Gemini Robotics 2 Controls a Humanoid From Feet to Fingertips — Capital Is Already Chasing the Same Bet

Google DeepMind introduced three robotics models this week: Gemini Robotics 2, a vision-language-action model controlling full humanoid robots down to 22-degree-of-freedom hand dexterity; Gemini Robotics ER 2, an embodied-reasoning model for multi-step task planning and multi-robot collaboration; and a lightweight on-device model that adapts to new robot bodies from just a few hours of demonstration data. On an Apollo 2 humanoid fitted with Inspire hands, DeepMind reports success rates of 68.4% picking objects from a table, 45.7% from the floor, and 76.3% from a shelf, with dexterous manipulation using parallel grippers reaching 74.2–89.6% depending on the task. The release also introduces ASIMOV-Agentic, a new safety benchmark measuring whether an embodied reasoning agent will refuse an unsafe tool call.

The model layer isn’t the only place this bet is being made. Days earlier, Travis Kalanick announced ATOMS, a new industrial-AI venture backed by a $1.7 billion investment led by Andreessen Horowitz, with Ben Horowitz joining the board. Kalanick frames it as the next stage of a “bits-to-atoms” career — after digitizing transportation with Uber and food-production logistics with CloudKitchens — now aimed at broad digitization of mining, construction, and food production through what he calls “atoms-based computers,” where manufacturing is the CPU and robotics is the execution layer. He states the ambition is to reshape multiple trillion-dollar industries within a decade.

Different layers, same underlying wager: DeepMind is publishing benchmark numbers on what a foundation model can make a humanoid robot’s hands do; a16z is putting $1.7 billion behind the industrial infrastructure to put that capability to work at scale. Neither is proof the wager pays off — robotics has a long history of benchmark success not translating cleanly to deployed reliability — but seeing model progress and venture capital move on the same timeline is a stronger signal than either alone. For engagements touching physical-operations AI strategy, this is the moment to start tracking robotics not as a research curiosity but as a category with real capital and real benchmarks attached.