On July 30, 2026, Google DeepMind announced Gemini Robotics 2, pushing its vision-language-action line a step further: robots no longer move only their upper bodies, but gain whole-body control from feet to fingertips. Carolina Parada, who leads robotics at DeepMind, positions this generation as “the intelligence layer powering the next generation of truly adaptable robots.”
One System, Three Models
Gemini Robotics 2 ships as three models. Gemini Robotics 2 itself is the flagship vision-language-action model (VLA), turning visual and language input directly into motor control; it can drive full humanoids and bi-arm robots. Gemini Robotics ER 2 is an embodied reasoning model (VLM) that acts as the robot’s high-level brain — communicating with people, understanding the physical world, and planning multi-step tasks that run several minutes. Gemini Robotics On-Device 2 is the efficiency-focused VLA built to run locally, offline, on robot hardware. ER 2 is available now on Google AI Studio and in private preview on the Gemini Enterprise Agent Platform; the other two models go to early-access partners first.
Whole-Body Control and Dexterous Hands
The clearest leap is whole-body work. Tell Apptronik’s Apollo 2 humanoid to “put the watering can into the green bin in the bottom shelf,” and it walks to the table, picks up the can, crosses to the shelves, and places it — previous generations controlled only the upper body for tabletop tasks. On dexterity, the same model checkpoint drives the 22-degree-of-freedom SharpaWave five-finger hand on Apollo 2 through delicate actions like tying knots or sealing a ziplock bag, and operates standard two-finger grippers on a Franka Duo platform for tight packing. DeepMind is candid about the limits: multi-finger dexterous manipulation remains challenging, and movement speed still has room to improve.
Agentic Reasoning and Robot Teamwork
ER 2 behaves like an agent: it observes the space, reasons through the steps, delegates execution to the VLA, and tracks progress — self-correcting when a step fails. This generation handles task sequences lasting several minutes and involving hundreds of decisions, and the model understands when tasks begin and end, pinpointing key events along the way. A new multi-robot collaboration feature lets different types of robots communicate and split workflows no single machine could finish alone; in the cluttered-room demo, robots team up to get the job done faster.
On-Device 2: New Bodies in Hours
Many robotics applications cannot depend on network connectivity. On-Device 2 inherits the “motion transfer” techniques from Gemini Robotics 1.5 and is natively multi-embodiment: for a new robot with drastically different shape, sensors, and degrees of freedom, adaptation typically takes a few hours and fewer than 200 examples, demonstrated across the Dexmate, SO101, and Trossen platforms. For hardware startups, the cost of “buy the model, fit the body” drops another notch.
Safety Design: The ASIMOV-Agentic Benchmark
As capability grows, the safety testing grows with it. DeepMind introduced ASIMOV-Agentic, a benchmark for agentic safety orchestration and uncertainty resolution: it measures whether the embodied reasoning agent refuses unsafe tool calls from a VLA, predicts whether a task is possible, and proactively requests human intervention when uncertain. ER 2 is also DeepMind’s safest robotics model to date on safety-constraint-following and human-proximity benchmarks — it detects when people are nearby, triggers safety tool calls, and brings the robot to a safe stop, a key requirement in collaborative safety standards. Details are in a safety technical report published alongside the release.
What It Means for Robotics Developers
Three signals stand out. First, the split between an ER model that plans and a VLA that executes is hardening into the standard architecture for robots, mirroring the planner-and-tools pattern in software agents; developers can evaluate each layer separately. Second, with multi-embodiment adaptation down to a few hours and about 200 examples, robot startups can concentrate on hardware design rather than training models from scratch. Third, by putting ER 2 into AI Studio and an enterprise agent platform, Google is pulling physical AI into the everyday developer toolchain. DeepMind is explicit about the destination — solving AGI in the physical world. That goal is distant, but this release moves humanoid robots from demo videos toward usable tasks.
Sources
- Gemini Robotics 2 brings whole body intelligence to robots — Google DeepMind
- Introducing Gemini Robotics ER 2 — Google
- Gemini Robotics — Google DeepMind
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.