# Gemini Robotics 2

> Source: https://aiwiki.ai/wiki/gemini_robotics_2
> Updated: 2026-07-31
> Categories: AI Models, Embodied AI, Google DeepMind, Multimodal AI, Robotics
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/)
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution to "AI Wiki (aiwiki.ai)".

**Gemini Robotics 2** is a family of three robotics models announced by [Google DeepMind](/wiki/google_deepmind) on July 30, 2026: a [vision-language-action](/wiki/vision_language_action_model) (VLA) model of the same name, an embodied reasoning model called Gemini Robotics ER 2, and an efficiency-oriented on-robot model called Gemini Robotics On-Device 2.[1] It is the second numbered generation of the [Gemini Robotics](/wiki/gemini_robotics) line and the first in which a single learned policy drives an entire [humanoid robot](/wiki/humanoid_robot), which Google DeepMind describes as control "from feet to fingertips" rather than the upper-body, tabletop control of earlier releases.[1] The launch was written by Carolina Parada, who leads robotics at Google DeepMind, and was accompanied by two model cards, a safety technical report, and a new open safety benchmark called ASIMOV-Agentic.[1][3][4][5]

| Field | Value |
|---|---|
| Developer | [Google DeepMind](/wiki/google_deepmind) |
| Announced | July 30, 2026[1] |
| Models | Gemini Robotics 2 (VLA), Gemini Robotics ER 2 (VLM), Gemini Robotics On-Device 2 (VLA)[1] |
| ER 2 base model | [Gemini 3.5 Flash](/wiki/gemini_3_5_flash)[3] |
| On-Device 2 base | Google's on-device [Gemma](/wiki/gemma) models plus Gemini Robotics 1.5 technology[4] |
| ER 2 context | 131,072 input tokens, 65,536 output tokens[9] |
| ER 2 endpoints | `gemini-robotics-er-2-preview`, `gemini-robotics-er-2-streaming-preview`[9] |
| Demonstrated hardware | [Apptronik](/wiki/apptronik) [Apollo 2](/wiki/apollo_2) ([SharpaWave](/wiki/sharpa) and Inspire hands), Franka Duo with Robotiq gripper, Dexmate, SO101, Trossen[1] |
| Access | ER 2 public preview; VLA and On-Device 2 by early-access partner program[1][7][8] |
| Documentation | Two model cards and a safety technical report; no general technical report published at launch[1][3][4][5] |

## Background: from tabletop arms to whole bodies

Google DeepMind has shipped a Gemini Robotics release roughly every three to four months since March 2025, with one longer six-month gap between Gemini Robotics 1.5 and ER 1.6, and each one has widened either the model's reasoning or the set of bodies it can drive. The pattern up to this point was that the embodied reasoning model reached public developer channels first, while the action model stayed with a small set of partners.[16][18][19]

| Date | Release | What it added |
|---|---|---|
| March 12, 2025 | Gemini Robotics and Gemini Robotics-ER | First VLA and embodied reasoning models in the family, both built on [Gemini](/wiki/gemini) 2.0; [Apptronik](/wiki/apptronik) named as humanoid partner[16] |
| June 24, 2025 | Gemini Robotics On-Device | First VLA in the family optimized to run locally on the robot and the first offered for fine-tuning, plus a Gemini Robotics SDK with a [MuJoCo](/wiki/mujoco) simulator and adaptation from as few as 50 to 100 demonstrations[17] |
| September 25, 2025 | Gemini Robotics 1.5 and Gemini Robotics-ER 1.5 | Two-model agentic stack; VLA thinks in natural language before acting; motion transfer across embodiments; ER 1.5 reached the [Gemini API](/wiki/gemini_api)[18] |
| April 14, 2026 | Gemini Robotics-ER 1.6 | Better spatial reasoning, multi-view understanding, and instrument reading, developed with [Boston Dynamics](/wiki/boston_dynamics) for facility inspection on Spot[19] |
| July 30, 2026 | Gemini Robotics 2 family | Whole-body humanoid control, multi-finger dexterity, video progress tracking, multi-robot collaboration, ASIMOV-Agentic[1][2] |

The gap Gemini Robotics 2 targets is the one between a robot that walks and a robot that manipulates. Earlier models in the family were trained and evaluated on [ALOHA](/wiki/aloha) 2 bi-arm platforms, a bi-arm Franka, and the upper body of Apptronik's Apollo humanoid.[16][17][18] Legs were somebody else's problem, usually a separate locomotion controller that handed control to a manipulation policy at fixed points.

## What "whole-body control" means here

Google DeepMind's product page for the VLA describes the new capability as controlling entire humanoid bodies while dynamically balancing the robot's center of gravity, so that a robot can step, squat, and bend instead of reaching from a fixed stance.[6] In the launch demonstration, Apollo 2 is asked to put a watering can into the green bin on the bottom shelf. It walks to the table, picks up the can, takes a few steps to the shelves, and places it.[1]

In plain terms: reaching for something on the floor changes where your weight is. A policy that only commands arms has to assume the base stays put, so somebody else has to decide when to walk, when to crouch, and how to keep the robot from tipping while the arms move a load around. Putting legs, torso, arms, and fingers under one policy means the model can trade off balance against reach inside a single decision rather than across a handoff between subsystems. Tech Times characterized this as collapsing the traditional separation between locomotion and manipulation controllers into one end-to-end policy, which is the standard framing for the problem the research community calls loco-manipulation.[22]

Google DeepMind is candid that speed is not solved. The launch post says the robots "have more to advance in movement speed" while calling whole-body coordination an important step toward harder real-world tasks.[1]

### One checkpoint, several bodies

The performance figure published with the launch shows a single model checkpoint driving three different embodiments: Apollo 2 with SharpaWave hands, Apollo 2 with Inspire hands, and a Franka Duo with a Robotiq gripper.[1] This matters more than it sounds. Cross-embodiment transfer has historically required per-robot specialization, and Google DeepMind first reported motion transfer across bodies with Gemini Robotics 1.5, where skills demonstrated only on ALOHA 2 also worked on Apollo and the bi-arm Franka.[18] Gemini Robotics 2 extends the idea to different end effectors on the same robot, so the same weights drive a five-fingered hand and a two-fingered parallel gripper.

The five-fingered hand in question is the SharpaWave, which Google DeepMind describes as having 22 degrees of freedom.[1] Fine manipulation with a [dexterous hand](/wiki/dexterous_hand) of that complexity is a different control problem from closing a gripper, and the published results reflect that.

### Fast adaptation on device

Gemini Robotics On-Device 2 inherits the motion transfer techniques from Gemini Robotics 1.5 and is natively multi-embodiment. Google DeepMind reports adapting it to new bi-arm embodiments in a few hours, typically with fewer than 200 examples, including bodies with substantially different shapes, sensors, and degrees of freedom. The platforms shown are Dexmate, SO101, and Trossen.[1][8] Its model card describes it as built on Google's on-device [Gemma](/wiki/gemma) models plus Gemini Robotics 1.5 technology, taking text, images, and robot proprioception as input and emitting robot actions as numerical values.[4]

## The reasoning and action loop

Gemini Robotics ER 2 is the high-level layer. It is a vision-language model, not an action model: it observes the scene, plans the steps, calls the VLA to execute each one, and tracks progress until the task is done, which lets the system run multi-step tasks lasting several minutes, self-correct after a failed step, and generalize to situations it has not seen.[1][15] Developers build this loop by declaring low-level control interfaces, such as VLA models or navigation APIs, as tools and streaming video, audio, or text into the model.[2] The model supports native [function calling](/wiki/function_calling), code execution, structured output, URL context, and grounding with [Google Search](/wiki/google_search).[7][9]

Google DeepMind reports that ER 2 integrates with the Gemini Live API through a bidirectional streaming endpoint so that it reasons about the next step while the current one is still executing, which the company says removes the stop-and-think pauses of earlier setups.[2] A separate streaming endpoint, `gemini-robotics-er-2-streaming-preview`, exists for exactly that use.[9]

Two capabilities are new in this generation and both are about time rather than space. Progress classification assigns each frame of a video feed to one of five completion bands, from 0 to 20 percent through 80 to 100 percent, so a robot knows roughly how far along it is. Moment finding identifies the exact frame where a critical event happens, such as when to stop pouring.[2]

### Multi-robot collaboration

The launch also introduced coordination across robots. Google DeepMind's framing is that different machine types have different strengths, so a wheeled rover and a humanoid can divide a workflow that neither could finish alone, communicating through a shared semantic understanding to hand off sub-tasks.[2][7] The demonstration pairs an Apptronik Apollo 2 humanoid with a Franka F3 Duo arm.[2] [Demis Hassabis](/wiki/demis_hassabis), chief executive of Google DeepMind, highlighted the same two capabilities in a post on X, singling out tying delicate knots and robots teaming up on complex workflows.[14]

## Reported results

All figures below are Google DeepMind's own, taken from the charts published with the launch post. The company notes that each whole-body and gripper bar is an average success rate across multiple tasks within a skill category, while the multi-finger figures are per-task, and that the same model checkpoint produced all of them.[1]

| Skill category | Platform | Task | Success rate |
|---|---|---|---|
| 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% |
| General whole-body manipulation | Apollo 2, Inspire hands | Pick up from shelf | 76.3% |
| Multi-finger dexterity | Apollo 2, SharpaWave hands | Screw bulb | 36% |
| Multi-finger dexterity | Apollo 2, SharpaWave hands | Unscrew bulb | 92% |
| Multi-finger dexterity | Apollo 2, SharpaWave hands | Tie trash bag | 44% |
| Multi-finger dexterity | Apollo 2, SharpaWave hands | Dustpan | 32% |
| Multi-finger dexterity | Apollo 2, SharpaWave hands | Ziplock | 40% |
| Gripper dexterity | Franka Duo | General pick and place | 74.2% |
| Gripper dexterity | Franka Duo | Diverse tool kitting | 78.9% |
| Gripper dexterity | Franka Duo | Precise insertion | 89.6% |

Google DeepMind's own caption on the chart says that while the model reaches medium to high success on whole-body and gripper-based tasks, multi-finger dexterous manipulation remains challenging.[1] The 56-point spread between unscrewing a bulb (92%) and screwing one in (36%) is the clearest signal in the package: the model can undo a threaded fitting far more reliably than it can start one.

For the reasoning model, Google DeepMind reports 57.4% accuracy on the five-band progress classification task and 91.3% accuracy on moment finding with a mean absolute distance of 0.96 seconds, and says ER 2 delivers that at a fraction of the compute cost and four times the execution speed of much larger model categories.[2] The company also reports that ER 2 beats ER 1.6 on tool orchestration across three control modes (a real VLA, a simulated VLA, and a human teleoperator) and achieves the highest accuracy across its core spatial capabilities, including success detection from image and video, embodied reasoning question answering, and generalized instrument reading tested across ten instrument types.[2]

## Safety and the ASIMOV-Agentic benchmark

Google DeepMind published a separate safety technical report dated July 29, 2026, and released the underlying benchmark as an open dataset on [Hugging Face](/wiki/hugging_face) at `google/asimov_agentic`.[5][11] ASIMOV-Agentic extends the earlier ASIMOV semantic safety benchmarks to agentic behavior: it asks whether the reasoning agent can refuse unsafe tool calls to the VLA, shield the VLA from tasks outside its training distribution, trigger protective stops during execution, and ask a human for help instead of guessing.[5]

| Component | What it measures | Reported result |
|---|---|---|
| Unsafe task refusal and constraint following | Classification accuracy across text, pointing, bounding box, and tool-use modalities | All frontier models tested scored at or above 96.0% in the text modality, with much wider variance once constraints had to be grounded in coordinates or tool calls[5] |
| Proactive human safety monitoring | Distance to nearest human from stereo images; whether to call `robot_stop()` | Top models estimated distance with 0.35 to 0.55 m mean absolute error and detected 1-meter breaches with roughly 79% to 93% accuracy[5] |
| Safety tool calling | Handling mid-task safety messages by severity | ER 2, Claude Opus 4.8, and GPT 5.5 all reached 100%[5] |
| VLA feasibility awareness | Whether the agent rejects tasks the VLA cannot do | ER 2 rose from 62.0% with no awareness of the VLA's training distribution to 95.8% at the most detailed tier[5] |
| Instruction ambiguity | Asking for clarification on under-specified commands, across a ten-class taxonomy | Reported as a helpfulness versus safety trade-off rather than a single score[5] |
| Obfuscated instrument reading | Refusing to hallucinate a reading from a damaged or badly lit gauge | Frontier models tended to prioritize avoiding hallucination[5] |

The most useful finding in the report is a negative one. In the agentic human-proximity evaluation, models that keep unnecessary stops below 5 percent miss more than 40 percent of genuine proximity hazards, while models that cut misses to the 10 to 15 percent range stop the robot unnecessarily 15 to 25 percent of the time. No model tested sits in the desirable corner of that trade-off, which the authors read as a reason to keep deterministic low-level guardrails underneath the model.[5]

In physical tests, Apollo 2 performed an object-sorting task in a garage while a human actor walked into the workspace from various angles. Google DeepMind reports that in lab settings ER 2 flagged human presence with 99% accuracy at configurable 1-meter and 2-meter thresholds and the VLA transitioned to a safe pose 96% of the time, after which the robot resumed the task on its own once the person left.[5] The report is explicit that it does not evaluate the certified hardware, redundancy, and real-time guarantees that a compliant deployment under standards such as ISO 10218:2025 would require, and points to ISO 13482 and the draft ISO 25785-1 as the separate track governing dynamically stable humanoids.[5] The ER 2 model card separately forbids use in safety-critical settings including healthcare and transportation.[3]

## Availability, pricing, and partners

Access is tiered by model, and only the reasoning model is publicly available.

| Model | Status | Route |
|---|---|---|
| Gemini Robotics ER 2 | Public preview | [Google AI Studio](/wiki/google_ai_studio) and the Gemini API; private preview on the Gemini Enterprise Agent Platform[1][7] |
| Gemini Robotics 2 (VLA) | Private preview | Waitlist for early access; Trusted Tester Program[1][6] |
| Gemini Robotics On-Device 2 | Trusted testers only | Trusted Tester Program signup[4][8] |

Google's published pricing puts `gemini-robotics-er-2-preview` at 2.00 US dollars per million input tokens (text, image, video, or audio) and 10.00 dollars per million output tokens including thinking tokens on the paid tier, with a free tier and context caching at 0.20 dollars per million tokens plus 1.00 dollar per million tokens per hour of storage. The streaming endpoint carries the same token rates. That is double the ER 1.6 preview rate for text, image and video input (1.00 dollar) and for output (5.00 dollars), though ER 1.6 already charged 2.00 dollars for audio input.[10] Google's documentation tells ER 1.6 users to migrate and says the 1.6 endpoint will be shut down at the end of August.[9]

Google DeepMind thanks Apptronik, [Boston Dynamics](/wiki/boston_dynamics), and [Agile Robots](/wiki/agile_robots) as partners in the launch post.[1] Boston Dynamics is the deepest of the three publicly: the two companies announced an AI partnership at CES on January 5, 2026 to put Gemini Robotics foundation models on Boston Dynamics' new Atlas humanoid, with joint research at both companies.[23] The ER 2 developer post also shows a Spot quadruped fetching a snack under ER 2 orchestration of Spot's navigation and manipulation APIs, with example code published on GitHub.[2]

## Reception, limits, and open questions

Coverage on the day of the launch focused on the same two things: that a single policy now drives legs as well as fingers, and that the published success rates are a long way from deployable. SiliconANGLE described ER 2 as the highlight of the release and emphasized progress tracking that lets a robot resume from the last correctly completed step instead of restarting a chore.[21] Tech Times treated the spread across the fine-motor tasks as the most important figure in the benchmark package, on the grounds that it draws the boundary between what a humanoid does reliably and what is still research.[22]

Engadget reported that Google says the launch video is real-time footage of fully autonomous robots, while noting that these are not general-purpose machines: each task in the video was specifically trained using a mix of human [teleoperation](/wiki/teleoperation), video examples, and simulations. The same piece contrasts the autonomy claim with an earlier high-profile [Tesla Optimus](/wiki/tesla_optimus) demonstration that was later shown to have leaned on teleoperators, and relays a comment Parada made to Wired about the safety problem ahead: "There's a lot of uncertainty that will show up, and so you want to be able to understand the safety question more deeply."[20]

Several limits are stated by Google DeepMind itself rather than by critics. Movement speed is called out in the launch post as needing more work.[1] The On-Device 2 model card says the model is limited in generalizing to out-of-distribution tasks and in controlling high-degree-of-freedom robots, and that because it was evaluated mainly on standing bi-arm manipulation, risks tied to mobile platforms or [whole-body control](/wiki/whole_body_control) are outside its evaluated scope.[4] No general technical report or peer-reviewable paper accompanied the family at launch, so the training recipe, data scale, control frequency, and evaluation protocol behind the success rates are not public in the way the March 2025 technical report documented the first generation.

## Competitive landscape

Gemini Robotics 2 lands in a field where several groups are building generalist policies for [robot manipulation](/wiki/robot_manipulation) and [physical AI](/wiki/physical_ai), each with a different theory of distribution.

[Figure AI](/wiki/figure_helix) reached full upper-body high-rate control first: Helix, announced February 20, 2025, was described as the first VLA to output high-rate continuous control of the entire humanoid upper body including wrists, torso, head, and individual fingers, and also as the first VLA to operate on two robots at once on a shared long-horizon task. Figure made a third precedence claim at the same time, that Helix was the first VLA to run entirely onboard embedded low-power GPUs.[24] That last point is worth holding next to Google DeepMind's multi-robot announcement: coordinating two identical robots under one policy is not new, but ER 2's version coordinates different robot types through a reasoning layer that allocates sub-tasks.[2][24]

NVIDIA has taken the opposite distribution route. Isaac GR00T N1, announced March 18, 2025, was released as an open, fully customizable humanoid foundation model with a dual-system architecture, alongside synthetic data blueprints and the Newton physics engine developed with Google DeepMind and Disney Research.[25] Where [Isaac GR00T](/wiki/isaac_gr00t) is meant to be post-trained by anyone, Gemini Robotics 2's action models remain behind a partner program, and only the reasoning layer is on a public API.

Other groups pursuing generalist robot policies include [Physical Intelligence](/wiki/physical_intelligence) and [Skild AI](/wiki/skild_ai), and vertically integrated programs such as Tesla's Optimus stack keep their models tied to their own hardware. The strategic distinction for Google DeepMind is that it builds no robots. Its bet is that the intelligence layer generalizes across [robot foundation model](/wiki/robot_foundation_model) customers, which is why the single-checkpoint result across Apollo 2, two different hand types, and a Franka gripper is doing more argumentative work in this release than any individual success rate.

Whether that bet holds depends on things this launch does not settle: how much [imitation learning](/wiki/imitation_learning) data each new embodiment actually needs beyond the reported few hours and 200 examples, how [robot learning](/wiki/robot_learning) results in a lab garage transfer to a working facility, and whether the multi-finger numbers move enough to make a 22-degree-of-freedom hand worth its cost compared with a gripper that already inserts parts at 89.6%.

## See also

- [Gemini Robotics](/wiki/gemini_robotics)
- [Vision-language-action model](/wiki/vision_language_action_model)
- [Embodied AI](/wiki/embodied_ai)
- [Robot foundation model](/wiki/robot_foundation_model)
- [Humanoid robot hands](/wiki/humanoid_robot_hands)
- [Robot safety](/wiki/robot_safety)
- [Apptronik](/wiki/apptronik)
- [Isaac GR00T](/wiki/isaac_gr00t)

## References

1. Parada, Carolina. "Gemini Robotics 2 brings whole body intelligence to robots." Google DeepMind, July 30, 2026. https://deepmind.google/blog/gemini-robotics-2-brings-whole-body-intelligence-to-robots/
2. Hansen, Steven and Peng Xu. "Introducing Gemini Robotics ER 2." Google (The Keyword), July 30, 2026. https://blog.google/innovation-and-ai/models-and-research/google-deepmind/gemini-robotics-er-2/
3. "Gemini Robotics ER 2 model card." Google DeepMind, published July 30, 2026. https://deepmind.google/models/model-cards/gemini-robotics-er-2/
4. "Gemini Robotics On-Device 2 model card." Google DeepMind, published July 30, 2026. https://deepmind.google/models/model-cards/gemini-robotics-on-device-2/
5. Gemini Robotics Team. "Gemini Robotics 2: Safety Evaluations." Google DeepMind, July 29, 2026 (PDF). https://storage.googleapis.com/deepmind-media/gemini-robotics/Gemini-Robotics-2-Safety.pdf
6. "Gemini Robotics 2." Google DeepMind model page, accessed July 31, 2026. https://deepmind.google/models/gemini-robotics/vla/
7. "Gemini Robotics ER 2." Google DeepMind model page, accessed July 31, 2026. https://deepmind.google/models/gemini-robotics/embodied-reasoning/
8. "Gemini Robotics On-Device 2." Google DeepMind model page, accessed July 31, 2026. https://deepmind.google/models/gemini-robotics/on-device/
9. "Gemini Robotics ER." Gemini API documentation, Google, last updated July 30, 2026. https://ai.google.dev/gemini-api/docs/robotics-overview
10. "Gemini Developer API pricing." Google, accessed July 31, 2026. https://ai.google.dev/gemini-api/docs/pricing
11. "google/asimov_agentic." Hugging Face dataset, Google, accessed July 31, 2026. https://huggingface.co/datasets/google/asimov_agentic
12. "Intelligent whole-body control with Gemini Robotics 2." Google DeepMind, YouTube, July 2026. https://www.youtube.com/watch?v=9MNLEAzA59o
13. Google DeepMind (@GoogleDeepMind). Post on X, July 30, 2026. https://x.com/GoogleDeepMind/status/2082844162928381956
14. Hassabis, Demis (@demishassabis). Post on X, July 30, 2026. https://x.com/demishassabis/status/2082883806323728416
15. Google AI (@GoogleAI). Post on X, July 30, 2026. https://x.com/GoogleAI/status/2082844740446253125
16. Parada, Carolina. "Gemini Robotics brings AI into the physical world." Google DeepMind, March 12, 2025. https://deepmind.google/blog/gemini-robotics-brings-ai-into-the-physical-world/
17. Parada, Carolina. "Gemini Robotics On-Device brings AI to local robotic devices." Google DeepMind, June 24, 2025. https://deepmind.google/blog/gemini-robotics-on-device-brings-ai-to-local-robotic-devices/
18. Parada, Carolina. "Gemini Robotics 1.5 brings AI agents into the physical world." Google DeepMind, September 25, 2025. https://deepmind.google/blog/gemini-robotics-15-brings-ai-agents-into-the-physical-world/
19. Graesser, Laura and Peng Xu. "Gemini Robotics-ER 1.6: Powering real-world robotics tasks through enhanced embodied reasoning." Google DeepMind, April 14, 2026. https://deepmind.google/blog/gemini-robotics-er-1-6/
20. "Google's new Gemini Robotics 2 platform allows for 'intelligent whole-body control'." Engadget, July 30, 2026. https://www.engadget.com/2227268/google-gemini-robotics-2-platform-intelligent-whole-body-control/
21. Deutscher, Maria. "Google DeepMind debuts Gemini Robotics 2 model series for humanoid robots." SiliconANGLE, July 30, 2026. https://siliconangle.com/2026/07/30/google-deepmind-debuts-gemini-robotics-2-model-series-humanoid-robots/
22. Hill, Terrence. "Gemini Robotics 2 Controls Full Humanoids: Legs, Torso, Arms, and Fingers Under One Policy." Tech Times, July 30, 2026. https://www.techtimes.com/articles/322309/20260730/gemini-robotics-2-controls-full-humanoids-legs-torso-arms-fingers-under-one-policy.htm
23. "Boston Dynamics and Google DeepMind Form New AI Partnership to Bring Foundational Intelligence to Humanoid Robots." Boston Dynamics, January 5, 2026. https://bostondynamics.com/blog/boston-dynamics-google-deepmind-form-new-ai-partnership/
24. "Helix: A Vision-Language-Action Model for Generalist Humanoid Control." Figure, February 20, 2025. https://www.figure.ai/news/helix
25. "NVIDIA Announces Isaac GR00T N1, the World's First Open Humanoid Robot Foundation Model, and Simulation Frameworks to Speed Robot Development." NVIDIA Newsroom, March 18, 2025. https://nvidianews.nvidia.com/news/nvidia-isaac-gr00t-n1-open-humanoid-robot-foundation-model-simulation-frameworks