Google DeepMind released Gemini Robotics 2 on July 30, 2026, a three-model family aimed at whole-body humanoid control, and published per-task success rates alongside it. The most revealing pair of numbers sits in multi-finger dexterity, where unscrewing a light bulb succeeds 92% of the time while screwing the same bulb in succeeds 36% of the time. ASAP works from Google DeepMind's official announcement as the primary source to map what this family can now do and where it remains stuck.

One model claiming everything from feet to fingertips

Google DeepMind frames whole-body control, spanning what it calls "feet to fingertips," as the defining advance in Gemini Robotics 2, and that framing separates it from Gemini Robotics 1.5, which centered on upper-body tabletop work. The stated goal is walking, crouching, reaching and manipulating objects simultaneously, along with task sequences that run several minutes and involve hundreds of decisions.

The family splits into three models. Gemini Robotics 2 is the Vision-Language-Action model that emits actions directly, Gemini Robotics ER 2 is the embodied reasoning model handling high-level judgment, and Gemini Robotics On-Device 2 runs locally on the robot itself. ER 2 is described as interpreting user instructions, observing the room, reasoning about required steps, coordinating with the VLA and tracking progress to completion. On-Device 2 operates without network latency or an internet connection.

Supported platforms include Apptronik Apollo 2, Franka Duo, Dexmate, SO101 and Trossen, with Boston Dynamics and Agile Robots systems also referenced. On-Device 2 inherits the motion transfer technique from Gemini Robotics 1.5, adapting to robots with drastically different shapes, sensors and degrees of freedom in a few hours and typically fewer than 200 examples.

The asymmetry hidden in the success-rate table

The most informative figures in Gemini Robotics 2's results are the gap between two operations on the same object. On Apollo fitted with SharpaWave hands, unscrewing a bulb succeeds 92% of the time and screwing one in succeeds 36% of the time, a 56-point spread. Tying a trash bag at 44%, ziplock at 40% and dustpan at 32% cluster in the same low band.

That asymmetry puts a number on a long-known property of robotic manipulation. Unscrewing requires repeated rotation until the object comes free, and error does not accumulate. Screwing in requires an alignment phase to find the thread, then force control after contact, and a failure at any stage fails the whole attempt. The 92 and the 36 reflect a difference in task structure, not a difference in capability.

The gripper table supports that reading. On Franka Duo, precise insertion scores 89.6%, higher than general pick and place at 74.2% or diverse tool kitting at 78.9%. Insertion itself is not the hard part; combining alignment and force control while holding an object in five fingers is. Google DeepMind states the same conclusion directly, noting that whole-body and gripper-based tasks reach medium to high success rates while multi-finger dexterous manipulation remains challenging.

What 45.7% says about the humanoid form factor

On whole-body manipulation, Apollo's success rate varies by object location: 76.3% from a shelf, 68.4% from a table and 45.7% from the floor. The ordering runs exactly inverse to how much the robot must change posture. A shelf needs only an arm extension from standing, a table needs modest torso flexion, and the floor demands a crouch or knee bend, a full postural transition.

That is precisely the case for and against the humanoid form. A wheeled robot arm has no floor row at all, because objects on the floor sit outside its workspace to begin with. The humanoid creates that row and takes on a new failure mode in exchange, balance during postural transition. A score of 45.7% says the trade is roughly half-completed, which is the reason to read whole-body control as entry rather than solution.

Applying the table to practical deployment is harder than it looks. The announcement does not state how many trials produced each figure or what counts as success per task. Success rates without trial counts carry no confidence intervals, so treating the roughly 8-point spread between 68.4% and 76.3% as a meaningful difference is unsupported. Widely separated values like 92% and 36% keep their direction regardless of trial count.

ER 2 now keeps watching the video

Gemini Robotics ER 2 operates on continuous video feeds rather than static snapshots, and it classifies task progress into five bands from 0-20% through 80-100% with 57.4% accuracy. On moment-finding, identifying critical events, it reaches 91.3% accuracy with a mean absolute distance of 0.96 seconds. Google DeepMind reports that ER 2 consistently outperforms ER 1.6 on tool orchestration across three control modes.

Those two numbers say different things. The 91.3% and 0.96 seconds indicate that identifying what happened and when, after the fact, is near practical quality. The 57.4% indicates that estimating how far along a task is, in real time, lands slightly better than half across five bands. Against a 20% random baseline for a five-way choice that is a clear gain, but it is too early to build a system that decides autonomously whether to retry on the strength of that number.

The more consequential structural change is that ER 2 connects to the Gemini Live API's bidirectional streaming endpoint, optimized for latency-sensitive tasks. Google DeepMind describes the result as letting the robot think about what comes next while performing actions, eliminating stop-and-think pauses. Given that perceived quality in robot demonstrations depends more on the gaps between motions than on the precision of any single motion, this change alters real-world impressions without appearing anywhere in the success-rate table.

ASIMOV-Agentic turns safety into a benchmark

Google DeepMind introduced the ASIMOV-Agentic benchmark with this release, and it measures the embodied reasoning agent's ability to refuse unsafe tool calls coming from a VLA. It also evaluates whether the agent can predict that a task is impossible and proactively request human intervention when uncertain. ER 2 is reported to outperform ER 1.6 and other frontier models on Safety Instruction Following and Human Proximity benchmarks, with halting a humanoid when humans are nearby given as the example.

The design direction deserves attention. ASIMOV-Agentic does not measure whether the robot moves safely; it measures whether the upper model declines the lower model's dangerous requests. Safety is redefined as a question of hierarchy between models rather than one of physical control. Once a dual structure is adopted, with the VLA generating actions and ER reviewing them, the reasoning is that real incidents originate at the interface between the two models rather than within any single motion.

No numerical ASIMOV-Agentic results accompany the announcement, however. Publishing manipulation success rates to one decimal place while handling a newly introduced safety benchmark through comparative prose alone is an asymmetry worth noting. Verifiable safety claims require figures and evaluation conditions in the same format as the rest.

What is usable now and what is not

Only one model in the Gemini Robotics family is generally accessible today: ER 2, available through the Gemini API and Google AI Studio, with a private preview on the Gemini Enterprise Agent Platform. The VLA model that emits actions and On-Device 2 go to early-access partners only, and everyone else must apply through the Trusted Tester Program.

That distribution pattern defines the character of this release. The reasoning model is open and the acting model is closed, so there is no route for outsiders to independently reproduce figures like 68.4% or 92% for now. Robotics benchmarks vary sharply with hardware and environment setup, which makes third-party verification unusually valuable here, and that path is blocked. The published numbers should be treated as Google DeepMind self-reporting.

Two questions matter most from a Korean industry standpoint. The first is whether the motion transfer claim, adaptation to a new embodiment from fewer than 200 examples, holds on the non-standard equipment common in domestic manufacturing. The second is whether teams can layer only the high-level reasoning tier onto their own hardware, exploiting the fact that ER 2 is exposed via API. Using ER 2's video understanding and orchestration without VLA access is the one path in this release that can be attempted immediately.

Source: Google DeepMind official announcements for Gemini Robotics 2 and Gemini Robotics ER 2 (July 30, 2026), compiled by ASAP