For most of the modern robotics era, the machines that shared our factories and warehouses were, in a quiet way, still frozen. They could weld a car door a hundred thousand times without complaint, but ask them to pick a watering can off an unfamiliar shelf and set it in a bin one room over, and the whole enterprise fell apart. On July 30, 2026, Google DeepMind released Gemini Robotics 2, an intelligence layer that is aimed squarely at that gap. Rather than a single model, it arrives as a family of three, and together they push robot autonomy past the table-top manipulation that defined the previous generation and into whole-body control, fine-fingered dexterity, and even cooperation between different kinds of robots.
The most cinematic demonstration involves Apptronik's Apollo 2 humanoid. Given the plainspoken instruction to "put the watering can into the green bin in the bottom shelf," Apollo walks to a table, picks up the can, takes a few steps toward a set of shelves, and places the object where it was asked. Earlier Gemini Robotics models could only command a humanoid's upper body for tasks performed while standing still; this one coordinates motion from, in DeepMind's phrase, feet to fingertips. The suite itself divides the labor deliberately. Gemini Robotics 2 is the vision-language-action model that turns what the robot sees and hears into motor commands. Gemini Robotics ER 2, an embodied-reasoning model built on Gemini 3.5 Flash, serves as the high-level brain that plans multi-step tasks stretching several minutes and hands motor execution off to the action model as if calling a tool. Gemini Robotics On-Device 2 is the efficient variant meant to run locally on the robot itself, for situations where network latency or a dropped connection is not an option.
DeepMind is unusually candid about the parts that still do not work well. A single model checkpoint drove three different bodies — Apollo 2 with five-fingered SharpaWave hands, Apollo 2 with Inspire hands, and a Franka Duo platform with a two-fingered gripper — and the reported success rates lay bare where the frontier actually sits. General whole-body picking ranged from 76.3% off a shelf down to 45.7% off the floor. Multi-finger dexterity was more sobering still: unscrewing a light bulb hit 92%, but tying a trash bag landed at 44%, sealing a ziplock bag at 40%, and sweeping with a dustpan at just 32%. The two-fingered gripper, by contrast, reached 89.6% on precise insertion. The lesson embedded in those numbers is that the closer a task gets to the delicate improvisation of a human hand, the further today's models have to go.
Why It Matters
It is tempting to read a humanoid tidying a shelf as a party trick, but the significance lies underneath the demo. The single hardest problem in robotics has never been building a capable body; it has been writing software general enough that skills learned on one machine survive the move to another. Gemini Robotics 2 attacks that directly. One checkpoint controlling a walking humanoid, a pair of dexterous hands, and a bench-mounted industrial arm is a claim about transfer, not about any one robot. If that claim holds up outside curated demonstrations, the economics of automation begin to shift, because the expensive part of a deployment stops being the hardware and starts being the intelligence that can be copied.
This is why the release lands with particular weight for manufacturing and logistics, the domains where repetitive physical labor is most abundant and most costly. A model that adapts to a new dual-arm platform in a few hours, using fewer than two hundred examples, points toward a future in which a warehouse operator does not commission a bespoke control system for every station but instead fine-tunes a shared foundation. DeepMind's own framing — that most robots today are pre-programmed or tele-operated for narrow, repetitive sequences that do not adapt to unpredictable environments — is really a description of the ceiling that industrial automation has pressed against for decades. Whole-body control and cross-embodiment learning are the first credible tools for lifting it, even if the dexterity numbers say the lifting will be gradual rather than sudden.
The Reaction
Within the research and developer community, the response has been a mixture of genuine enthusiasm and careful qualification, which is roughly the correct posture. The most discussed technical contribution was not the humanoid footage at all but ER 2's grasp of time. The model was measured on a task that rarely gets benchmarked: knowing when a job is actually finished. On progress classification, sorting each frame of a video into one of five completion bands, ER 2 reached 57.4% accuracy, which DeepMind reports as ahead of both previous-generation and competing frontier models. On "moment finding" — identifying the exact instant a critical event occurs, such as when to stop pouring coffee — it hit 91.3% accuracy while running several times faster than much larger models.
Developers also noted what DeepMind chose to open and what it kept closed. Gemini Robotics ER 2 is available in public preview through the Gemini API and Google AI Studio, so anyone can begin building high-level planning on top of it. The action model, though, remains limited to early-access partners, and the on-device version to trusted testers. Alongside the models, DeepMind published ASIMOV-Agentic, a new safety benchmark, on Hugging Face under a permissive CC-BY-4.0 license. That combination — an open reasoning layer and a shared safety yardstick, wrapped around gated motor-control weights — reads as a deliberate attempt to invite scrutiny of the thinking while holding the direct-actuation capability closer. It is a compromise, and like most compromises it satisfies no one completely, but it reflects a field learning to be careful in public.
What Comes Next
The near-term roadmap is written plainly in the limitations. DeepMind states that its robots still have progress to make in movement speed, and the dexterity gap between a 92% bulb unscrew and a 32% dustpan sweep is the kind of unevenness that gets closed one stubborn task at a time. On-Device 2 is described as limited in generalizing to out-of-distribution tasks and in controlling robots with many degrees of freedom, which is another way of saying the hardest, highest-dimensional problems are precisely the ones still waiting.
What is more interesting is the architecture the release quietly normalizes. By treating a vision-language-action model as a callable tool that a reasoning model orchestrates — and by demonstrating ER 2 driving a Boston Dynamics Spot through its navigation and manipulation APIs — DeepMind is sketching a world where a single high-level brain conducts a heterogeneous fleet: a wheeled rover for smooth indoor floors, a humanoid for uneven terrain, each handing subtasks to the other through shared semantic understanding. That is a meaningfully different picture from the one-robot-one-program paradigm, and it suggests the next competitive battles will be fought less over individual robot bodies than over the reasoning layer that can command any of them.
Closing Thoughts
There is a temptation, whenever a robot does something that looks fluid on video, to narrate it as the arrival of the future. The more honest reading of Gemini Robotics 2 is that it marks a shift in where the difficulty lives. For years the constraint was generality — the same software could not survive a change of body or a change of room. That constraint is loosening, and what remains exposed underneath is the older, more human problem of touch: the confidence with which a hand ties a knot or seals a bag without crushing it. Those are the tasks a two-year-old masters without instruction and a state-of-the-art model still fails more than half the time.
Perhaps that is the quietly humbling part of this moment. We have taught a machine to plan a several-minute task, to know when it has finished, to walk across a room and adapt itself to a body it has never worn before — and it is the small, dextrous, unglamorous gestures that keep it honest. The trajectory is unmistakable and the pace is real, but so is the distance left to travel. Gemini Robotics 2 is best understood not as a destination but as a clear, well-documented map of the ground that still lies ahead, published by a group willing to show its failures alongside its wins. In a field prone to spectacle, that candor may be the most useful thing it offers.
한글 요약
구글 딥마인드가 2026년 7월 30일 로봇 지능 계층 '제미나이 로보틱스 2'를 공개했다. 단일 모델이 아니라 세 종류로 구성되는데, 시각·언어·행동을 운동 명령으로 바꾸는 VLA 모델, 여러 단계 작업을 계획하는 상위 두뇌 격의 ER 2(제미나이 3.5 플래시 기반), 그리고 로봇에서 자체 구동되는 온디바이스 2로 나뉜다. 이전 세대가 상반신만 다뤘던 것과 달리, 앱트로닉의 아폴로 2 휴머노이드가 "물뿌리개를 아래 선반 초록색 통에 넣어라"는 지시를 받고 걸어가 집어 옮기는 전신 제어를 처음으로 선보였다.
주목할 점은 하나의 체크포인트가 서로 다른 세 로봇 몸체를 제어했다는 것으로, 이는 '전이(transfer)'에 관한 주장이다. 로봇공학의 오랜 난제는 뛰어난 몸체를 만드는 것보다 한 기계에서 배운 기술이 다른 기계로 옮겨가도록 하는 일이었다. 새 양팔 플랫폼에 수백 개 미만의 예시와 몇 시간이면 적응한다는 점은 제조·물류 자동화의 경제성이 하드웨어가 아니라 복제 가능한 지능 쪽으로 이동할 수 있음을 시사한다. 다만 딥마인드는 한계도 솔직하게 공개했다. 전구를 푸는 작업은 92% 성공했지만 쓰레기봉투 묶기 44%, 지퍼백 밀봉 40%, 쓰레받기 32%로, 사람 손의 섬세한 즉흥성에 가까울수록 성공률이 급격히 떨어졌다.
커뮤니티 반응은 열광과 신중한 유보가 뒤섞였다. 가장 많이 논의된 것은 휴머노이드 영상이 아니라 ER 2의 '시간 이해', 즉 작업이 언제 끝났는지 아는 능력이었다(순간 포착 정확도 91.3%). ER 2는 제미나이 API와 구글 AI 스튜디오에서 공개 프리뷰로 열렸지만 행동 모델과 온디바이스 모델은 각각 초기 파트너·신뢰 테스터로 제한됐고, 안전 벤치마크 ASIMOV-Agentic은 허깅페이스에 CC-BY-4.0으로 공개됐다. 성공 사례와 실패 수치를 함께 내놓은 이 투명함이야말로, 화려함에 치우치기 쉬운 분야에서 이번 발표가 남긴 가장 값진 부분일지도 모른다.
참고: Google DeepMind — Gemini Robotics 2, MarkTechPost, The Robot Report