Google DeepMind has released Gemini Robotics 2, the intelligence layer for its next generation of robots. The release moves the stack past table-top manipulation into whole body control, five finger dexterity and multi robot teamwork. It ships as three separate models with three different access tiers.
Most robots today are pre-programmed or tele-operated for narrow, repetitive task sequences. They do not adapt to unpredictable environments, and skills rarely transfer between robot bodies. Gemini Robotics 2 targets all three limits at once.
TL;DR
- Three models ship together: a VLA, an embodied reasoning VLM, and an on-device VLA.
- One checkpoint drives Apollo 2 with two different hands plus a Franka Duo gripper.
- Gemini Robotics ER 2 is public preview; the VLA and on-device models stay gated.
- Multi-finger dexterity is still the weak axis, ranging from 32% to 92%.
- ASIMOV-Agentic, a new safety benchmark, is on Hugging Face under CC-BY-4.0.
3 Models and What They Do
Gemini Robotics 2
Gemini Robotics 2 is the vision-language-action (VLA) model. It converts vision and language input into motor control. It can drive full humanoids from feet to fingertips, and other bi-arm robots. It also handles dexterous manipulation on both multi-finger hands and parallel grippers.
Gemini Robotics ER 2
Gemini Robotics ER 2 is the embodied reasoning (ER) model. It is a vision language model that acts as the high level brain. It communicates with humans, understands the physical world, and plans multi-step tasks lasting several minutes. Per its model card, ER 2 is based on Gemini 3.5 Flash. It accepts interleaved text, image, video and audio with a context window of up to 128k, and emits text up to 64K tokens.
Gemini Robotics On-Device 2
Gemini Robotics On-Device 2 is the efficient VLA optimized to run locally on the robot. Its model card states it is built on Gemini Robotics 1.5 technology and Google’s on-device Gemma models. Inputs are text, images and robot proprioception as numerical values. Outputs are robot actions as numerical values.
The division of labor matters for system design. ER 2 plans and tracks the task, then hands motor execution to a VLA declared as a tool. Developers register low level control interfaces, such as VLA models or navigation APIs, as callable tools. They then stream multimodal video, audio or text directly into the model.
Whole body control on Apptronik Apollo 2
Previous Gemini Robotics models controlled only the humanoid’s upper body for table-top tasks. Gemini Robotics 2 extends control to whole body motion for the first time.
The worked example uses Apptronik’s Apollo 2. Given the instruction “put the watering can into the green bin in the bottom shelf,” Apollo walks to the table and picks up the watering can. It then takes a few steps to the shelves and places the object at its destination.
Google DeepMind is direct about the remaining gap. It states that its robots have more to advance in movement speed.
Dexterity across hands and grippers
Gemini Robotics 2 can control the five-fingered, 22 degree-of-freedom SharpaWave hand on Apollo 2. Reported actions include tying knots and sealing a ziplock bag. The same model also operates standard two-fingered parallel grippers on a Franka Duo platform for tasks like tight packing.
The reported success rates come from one model checkpoint controlling three embodiments: Apollo 2 with SharpaWave hands, Apollo 2 with Inspire hands, and Franka Duo with a Robotiq gripper.
| Category | Embodiment | Task | Success |
|---|---|---|---|
| General whole body manipulation | Apollo 2 + Inspire hands | Pick up from shelf | 76.3% |
| General whole body manipulation | Apollo 2 + Inspire hands | Pick up from table | 68.4% |
| General whole body manipulation | Apollo 2 + Inspire hands | Pick up from floor | 45.7% |
| Multi-finger dexterity | Apollo 2 + Sharpa hands | Unscrew bulb | 92% |
| Multi-finger dexterity | Apollo 2 + Sharpa hands | Tie trash bag | 44% |
| Multi-finger dexterity | Apollo 2 + Sharpa hands | Ziplock | 40% |
| Multi-finger dexterity | Apollo 2 + Sharpa hands | Screw bulb | 36% |
| Multi-finger dexterity | Apollo 2 + Sharpa hands | Dustpan | 32% |
| Gripper dexterity | Franka Duo | Precise insertion tasks | 89.6% |
| Gripper dexterity | Franka Duo | Diverse tool kitting | 78.9% |
| Gripper dexterity | Franka Duo | General pick and place | 74.2% |
The developer X post focuses on a problem that rarely gets benchmarked: knowing when a task is actually done.
Progress classification:
Each frame in a video feed is assigned to one of five progress levels, from 0-20% through 80-100%. Gemini Robotics ER 2 reaches 57.4% accuracy on this task. Google DeepMind reports this outperforms previous generation models and competing frontier models.
Moment finding:
This measures whether the model can identify the exact frame where a critical event occurs. An example is the moment to stop pouring coffee into a cup. ER 2 reaches 91.3% accuracy with a 0.96 second mean absolute distance. Google DeepMind reports it competing closely with much larger model categories at 4x the execution speed.
Tool orchestration:
ER 2 is evaluated across three control modes: real VLA, sim VLA, and human tele-op. It outperforms Gemini Robotics ER 1.6 on all three.
ER 2 integrates with the Gemini Live API through a bidirectional streaming endpoint. The stated purpose is removing the stop-and-think pauses that break multi-step execution. It can also natively call tools like Google Search or any user-defined function.
Three spatial capabilities were upgraded. Success and failure detection now runs on raw video feeds instead of static snapshots, which catches mid-execution spills and slips. General instrument reading extends beyond circular dials to digital displays, linear scales, rulers and liquid thermometers, tested across 10 instrument types. Spatial visual question answering improves through Gemini’s multimodal advances.
Google DeepMind built a demo with Spot from Boston Dynamics, using ER 2 to orchestrate Spot navigation and manipulator APIs. Sample code is available in the robotics-samples repository.
Multi-Robot Collaboration
Gemini Robotics 2 introduces collaboration between different robot types. The reasoning is that no single robot fits every task. A wheeled rover suits indoor work, while a humanoid handles uneven terrain better.
Robots communicate through a shared semantic understanding to hand off subtasks. The demonstrated pairing is Apptronik’s Apollo 2 with a Franka F3 Duo.
On-Device 2: Adapting to New Robot Bodies
Gemini Robotics On-Device 2 targets applications that cannot rely on network latency or connectivity. It is natively multi-embodiment and inherits the motion transfer techniques from Gemini Robotics 1.5.
Google DeepMind reports adaptation to new bi-arm embodiments in a few hours, typically with fewer than 200 examples. This holds for embodiments with drastically different shapes, sensors and degrees of freedom. Demonstrated platforms include Dexmate, SO101 and Trossen.
The model card publishes a data scaling comparison against On-Device 1 on platforms introduced only during post-training:
| Platform | Model | Start | End |
|---|---|---|---|
| SO101 | On-Device 2 | 6.7% | 53.3% |
| SO101 | On-Device 1 | 0.0% | 6.7% |
| Dexmate | On-Device 2 | 24.4% | 75.6% |
| Dexmate | On-Device 1 | 13.3% | 33.3% |
The SO101 result is the sharper one. On-Device 1 barely leaves the floor, while On-Device 2 crosses 50%.
The model card also lists limits. On-Device 2 is limited in generalizing to out-of-distribution tasks and in controlling high-degree-of-freedom robots.
Availability
The AI Studio launch link uses the model string gemini-robotics-er-2-preview. Getting-started notebooks are in the robotics-samples repository.
Interactive explainer
‘;
}
return h;
}
var BENCH={
‘whole’:{t:’GENERAL WHOLE BODY MANIPULATION’,s:’Apptronik Apollo 2 · Inspire hands’,
r:[[‘Pick up from shelf’,76.3],[‘Pick up from table’,68.4],[‘Pick up from floor’,45.7]],
n:’Each bar is the average success rate over multiple tasks in that skill category. Error bars reported in the source chart are not reproduced here.’},
‘multi’:{t:’MULTI-FINGER DEXTERITY’,s:’Apollo 2 · SharpaWave hands (5-finger, 22 DoF)’,
r:[[‘Unscrew bulb’,92],[‘Tie trash bag’,44],[‘Ziplock’,40],[‘Screw bulb’,36],[‘Dustpan’,32]],
n:’These are individual task scores, not category averages. Google DeepMind states multi-finger dexterous manipulation remains challenging.’},
‘grip’:{t:’GRIPPER DEXTERITY’,s:’Franka Duo · Robotiq gripper’,
r:[[‘Precise insertion tasks’,89.6],[‘Diverse tool kitting’,78.9],[‘General pick and place’,74.2]],
n:’All three embodiments above were driven by the same Gemini Robotics 2 model checkpoint.’}
};
var SCALE={
‘SO101’:[[‘Gemini Robotics On-Device 2 — start’,6.7],[‘Gemini Robotics On-Device 2 — end’,53.3],[‘On-Device 1 — start’,0.0],[‘On-Device 1 — end’,6.7]],
‘Dexmate’:[[‘Gemini Robotics On-Device 2 — start’,24.4],[‘Gemini Robotics On-Device 2 — end’,75.6],[‘On-Device 1 — start’,13.3],[‘On-Device 1 — end’,33.3]]
};
function scaleBlock(k){
return ‘
ON-DEVICE 2 vs ON-DEVICE 1 — DATA SCALING ON A NEW EMBODIMENT
‘
+’
‘
+’‘
+’‘
+’
‘+bars(SCALE[k])
+’
Success rate as average hours of data per task increases, on platforms introduced only during post-training. Start = lowest reported data point, end = highest. Google DeepMind reports adaptation to a new bi-arm embodiment in a few hours, typically with fewer than 200 examples.
‘;
}
var VIEWS={
‘help’:function(){return ‘
GEMINI ROBOTICS 2 — INTERACTIVE EXPLAINER
‘
+’
Google DeepMind shipped three models on 30 July 2026. Pick a command:
‘
+’
‘
+’
| COMMAND | WHAT IT SHOWS |
|---|---|
| models | The three-model stack and what each one consumes and emits |
| bench –whole-body | Whole-body manipulation success rates on Apollo 2 |
| bench –multifinger | Five-finger SharpaWave hand task scores |
| bench –gripper | Two-finger gripper scores on Franka Duo |
| er2 –metrics | Progress classification, moment finding, orchestration |
| ondevice –scaling | On-Device 2 vs On-Device 1 on new embodiments |
| safety | ASIMOV-Agentic benchmark and usage restrictions |
| access | Who can use which model today |
Hover any bar for its exact value. Press Tab in the input to cycle commands.
‘;},
‘models’:function(){return ‘
THE THREE-MODEL STACK
‘
+’
+’
Gemini Robotics 2
Vision-language-action model.
In: vision + language
Out: motor control
Drives full humanoids feet-to-fingertips, plus bi-arm robots, hands and grippers.
‘
+’
Gemini Robotics ER 2
Embodied-reasoning VLM, the high-level brain.
Base: Gemini 3.5 Flash
In: text / image / video / audio, up to 128k context
Out: text, up to 64K tokens
‘
+’
Gemini Robotics On-Device 2
Efficient VLA that runs locally.
Built on Gemini Robotics 1.5 plus on-device Gemma models.
In: text, images, proprioception
Out: robot actions as numbers
‘
+’
‘
+’
ER 2 plans and tracks the task, then hands motor execution to a VLA declared as a tool. ER 2 can also call Google Search or any user-defined function natively, and streams over the Gemini Live API to avoid stop-and-think pauses.
‘;},
‘bench –whole-body’:function(){var b=BENCH.whole;return ‘
‘+b.t+’
‘+b.s+’
‘+bars(b.r)+’
‘+b.n+’
‘;},
‘bench –multifinger’:function(){var b=BENCH.multi;return ‘
‘+b.t+’
‘+b.s+’
‘+bars(b.r)+’
‘+b.n+’
‘;},
‘bench –gripper’:function(){var b=BENCH.grip;return ‘
‘+b.t+’
‘+b.s+’
‘+bars(b.r)+’
‘+b.n+’
‘;},
‘er2 –metrics’:function(){return ‘
GEMINI ROBOTICS ER 2 — REPORTED NUMBERS
‘
+bars([[‘Progress classification accuracy’,57.4],[‘Moment-finding accuracy’,91.3]])
+’
‘
+’
| METRIC | VALUE |
|---|---|
| Moment-finding mean absolute distance | 0.96 s |
| Execution speed vs much larger model categories | 4x |
| Progress buckets per video frame | 5 (0-20 … 80-100%) |
| Tool-orchestration control modes tested | real VLA, sim VLA, human tele-op |
| Instrument types in reading eval | 10 |
‘
+’
Google DeepMind reports ER 2 beating ER 1.6 on tool orchestration across all three control modes, and beating prior-generation and competing frontier models on progress classification. Success and failure detection now runs on raw video rather than static snapshots.
‘;},
‘ondevice –scaling’:function(){return scaleBlock(‘SO101’);},
‘safety’:function(){return ‘
SAFETY LAYER
‘
+’
| ITEM | DETAIL |
|---|---|
| New benchmark | ASIMOV-Agentic, released on Hugging Face under CC-BY-4.0 |
| What it tests | Refusing unsafe VLA tool calls, physical-constraint reasoning, feasibility prediction, uncertainty resolution |
| Human safety monitoring | Stereo-image proximity detection; default stop threshold 3.0 m |
| Gauge reading | Adversarial glare, blur and occlusion, plus expressing uncertainty when unreadable |
| Claimed result | ER 2 is Google DeepMind’s safest robotics model on Safety Instruction Following and Human Proximity |
‘
+’
The model card is explicit: do not use these models for safety-critical work, including healthcare, transportation, or any setting where a malfunction could foreseeably cause death, injury or property damage.
‘;},
‘access’:function(){return ‘
WHO CAN USE WHAT, TODAY
‘
+’
| MODEL | AVAILABILITY |
|---|---|
| Gemini Robotics ER 2 | Public preview — Gemini API, Google AI Studio; private preview on Gemini Enterprise Agent Platform |
| Gemini Robotics 2 (VLA) | Early-access partners only |
| Gemini Robotics On-Device 2 | Trusted Testers only |
‘
+’
‘
+’
Model string used by the AI Studio launch link:
‘
+’
> gemini-robotics-er-2-preview
‘
+’
Hardware seen in the release: Apptronik Apollo 2 with SharpaWave and Inspire hands, Franka Duo with a Robotiq gripper, Boston Dynamics Spot orchestrated through Spot APIs, and Dexmate, SO101 and Trossen platforms for on-device adaptation.
‘;}
};
function typeCmd(c,cb){
var d=el(‘
$
‘);
var t=d.querySelector(‘#gr2typ’),i=0;
(function tick(){
if(i<=c.length){t.textContent=c.slice(0,i);i++;scroll();setTimeout(tick,16);}
else{t.removeAttribute(‘id’);cb();}
})();
}
function run(c,fromChip){
if(busy)return;
c=(c||”).trim();
if(!c)return;
busy=true;
if(c===’clear’){out.innerHTML=”;mark(‘clear’);busy=false;boot();return;}
mark(c);
typeCmd(c,function(){
var load=el(‘
resolving…
‘);
setTimeout(function(){
load.parentNode.removeChild(load);
var fn=VIEWS[c];
if(!fn){
el(‘
command not found: ‘+c.replace(/ — type help
‘);
busy=false;scroll();return;
}
var block=el(fn());
trim();
anim(block);
bindTogs(block);
busy=false;scroll();
},260);
});
}
function bindTogs(scope){
var tbs=scope.querySelectorAll(‘.gr2-tb’);
for(var i=0;i