Gemini Robotics 2: What Evolved from the 1.x Line

01The three models

The three are not variants of one another. They occupy different layers of the same stack — deciding, moving, and running on the robot itself — and a working system uses more than one of them at a time.

Splitting the work this way is what DeepMind's own safety report calls a "system 2 / system 1" architecture: a deliberate planner on top, a fast reflexive controller underneath. It also explains why only one of the three is openly available to developers, which the access section below returns to.

02What changed since 1.x

Version 2 is the fifth release in a line that opened in March 2025, and the three lines have advanced on separate clocks ever since. Laid side by side, the jump from 1.x looks different in each of them.

Not everything was replaced. The two mechanisms 1.5 introduced are the ones version 2 pushes further: Motion Transfer, which carries a skill demonstrated on one robot body over to a different one, and an internal reasoning step in natural language before acting. Read that way, whole-body control and minutes-long planning are less new departures than those two ideas taken further.

The reasoning model has shipped three times without a matching action-model release, because it is the half developers can reach through an ordinary API. The name shifted with this release too: it was written Gemini Robotics-ER through 1.6, and Gemini Robotics ER 2 without the hyphen from version 2 onward.

03Whole-body control

Earlier Gemini Robotics models were, in practice, upper-body models. They moved arms and hands across a table, and anything outside that reach was somebody else's problem. Gemini Robotics 2 adds the legs and torso, and with them the robot's centre of gravity, so it can step, squat and bend toward a target instead of being limited to what already sits in front of it.

DeepMind's worked example is the instruction "put the watering can into the green bin in the bottom shelf." Apptronik's Apollo 2 humanoid walks to the table, picks the can up, walks to the shelves and places it, with the model handling both the navigation and the balance that reaching a bottom shelf requires.

The floor figure is the one worth sitting with. Reaching down for an object on the ground is the case that most needs whole-body control, and it is also where the published number is lowest, which is a fair summary of how new this capability is.

04Dexterity

The second addition is control of hands complex enough to be hard to program by hand. Gemini Robotics 2 drives Apollo 2 fitted with five-fingered SharpaWave hands carrying 22 degrees of freedom, and also drives ordinary two-fingered parallel grippers, such as a Robotiq gripper on a Franka Duo. That range matters commercially: most robots already deployed in industry have a simple gripper, not an anthropomorphic hand.

The task list DeepMind published is deliberately fiddly. Tying a knot in a bin bag, sealing a ziplock, sweeping into a dustpan and screwing a light bulb into a socket all need continuous force adjustment against an object that moves as you work on it.

Unscrewing a bulb and screwing one back in are nearly the same motion in reverse, and the gap between the two published rates is the widest in the set. Removal tolerates approximation, because the thread guides itself once it catches. Insertion does not: the bulb has to be aligned before any force is useful.

Two-fingered grippers score higher across the board on the tasks DeepMind reported for them, precise insertion included. A more capable hand, in other words, is not automatically a more reliable one.

05Multi-robot collaboration

The third addition belongs to ER 2 rather than to the action model. One reasoning model can now hold a plan across several robots at once, hand each of them the part of the job its body suits, and keep track of who has finished what. DeepMind's demonstration pairs an Apollo 2 humanoid with a Franka F3 Duo bi-arm manipulator; a separate build uses ER 2 to orchestrate Boston Dynamics' Spot through its navigation and manipulator APIs.

What makes this practical is a change in what ER 2 watches. It reads a continuous video feed rather than still frames, so it can judge when a step is actually finished, notice when one has failed, and re-plan mid-task. DeepMind describes tasks that run for several minutes across hundreds of decisions, a horizon that a model working from single snapshots cannot supervise.

Progress classification is the weakest of these in absolute terms, and it is also the one the multi-robot case depends on most: a coordinator that cannot reliably tell how far along a task is will hand out the next instruction at the wrong moment.

06Safety evaluation

Alongside the models, DeepMind published a safety report and released its benchmark, ASIMOV-Agentic, as a public dataset. The argument it makes is that physical safeguards such as emergency stops and speed limits remain necessary but are no longer sufficient, because a model that plans its own sub-tasks can walk into a hazard before any of those safeguards has anything to react to.

The results are uneven in an instructive way. Every frontier model tested, ER 2 included, scored 100% on reacting to structured safety messages, which is the part closest to conventional software. Judging whether a task is feasible in the first place depended heavily on how much the model had been told about the action model's training: accuracy ran from 62.0% with no such briefing to 95.8% with the most detailed one.

The report is also explicit about its own scope, and that matters more than any single figure in it. It evaluates high-level semantic reasoning only, and names three things a compliant physical deployment needs that it does not assess:

Scoring well here is not the same as being cleared to work next to people. That layer is separate engineering work, governed by standards such as ISO/TS 15066.

07Access

Only the reasoning model is generally available. The two action models remain behind an early-access programme, which is the same posture DeepMind took with every previous Gemini Robotics VLA.

The hardware side is broader than the headline humanoid. DeepMind names Apptronik, Boston Dynamics and Agile Robots as partners, and its reported evaluations span Apollo 2, the Franka Duo, and, for the on-device model, Dexmate, Trossen and SO-101 platforms, the last of which is a low-cost open-source arm.

08FAQ

09Sources