Gemini Robotics 2: Google DeepMind Brings Whole-Body Intelligence to Humanoid Robots

Google DeepMind announced Gemini Robotics 2 on July 30, a family of three models positioned as the "intelligence layer" for the next generation of robots. The headline claim is whole-body intelligence: for the first time, DeepMind says its vision-language-action model can control an entire humanoid robot from feet to fingertips, not just an arm on a tabletop. It also brings multi-finger dexterity, multi-robot collaboration, and a fast-adapting on-device model, and one of the three models is publicly testable right now.

The news hit the Hacker News front page with 575 points and 459 comments, the discussion split between impressed engineers and sharp-eyed skeptics.

What Google DeepMind Announced#

Gemini Robotics 2 is three models, not one:

• Gemini Robotics 2 - DeepMind's most advanced vision-language-action (VLA) model. It converts vision and language input directly into motor commands and can control full humanoids (feet to fingertips) plus bi-arm robots, with a claimed step up in dexterous manipulation on both multi-finger hands and two-finger grippers.
• Gemini Robotics ER 2 - the embodied reasoning (ER) model, a vision-language model that acts as the robot's high-level "brain." It communicates with humans, plans multi-step tasks lasting several minutes, tracks progress, and can now coordinate multiple robots working as a team.
• Gemini Robotics On-Device 2 - an efficiency-focused VLA that runs locally on the robot. DeepMind says it adapts to completely new robot embodiments in just a few hours, typically with fewer than 200 examples, inheriting the motion-transfer techniques from Gemini Robotics 1.5.

The numbers DeepMind shared are modest but concrete. On whole-body manipulation with an Apptronik Apollo 2 fitted with Inspire hands, success rates were 68.4% for picking up from a table, 45.7% from the floor, and 76.3% from a shelf. On multi-finger dexterity with Sharpa hands, results ran from 32% (dustpan) to 92% (unscrew bulb), with screw bulb at 36%, tie trash bag at 44%, and ziplock at 40%. Gripper dexterity on a Franka Duo was stronger: 74.2% pick-and-place, 78.9% tool kitting, and 89.6% precise insertion. DeepMind's own caption notes multi-finger dexterity "remains challenging."

On safety, the release introduces ASIMOV-Agentic, a new benchmark for agentic safety orchestration and uncertainty resolution, and DeepMind calls ER 2 its safest robotics model to date on safety-constraint-following and human-proximity benchmarks, with a companion safety technical report.

Partners include Apptronik, Boston Dynamics, and Agile Robots. The developer blog also shows a Boston Dynamics Spot fetching a snack on a natural-language command, orchestrated by ER 2 through Spot's APIs.

What HN Is Saying#

The Hacker News thread on the announcement post split into impressed engineers, hardware skeptics, and hands-on testers.

The insider take. DeepMind researcher canyon289 showed up to vouch for the lab: "I'm a researcher at Deepmind that contributed to these models. (And the opinions here are my own)" and called DeepMind "one of the few unique labs where you can move from large frontier models (Gemini), frontier open models (Gemma), robotics (what you see here), science (weather, biology, more)." The comment drew a long thread, including a good-natured nitpick from jauntywundrkind about AI2's breadth, which canyon289 accepted.

The actuator debate. Geee argued the real bottleneck is hardware, not AI: "There has been no innovation in robotic actuators since Honda's Asimo. There's just no way that someone wants a 80kg wobbling tin can in their home or workplace." siekmanj pushed back with specifics: "The torque density and price of actuators has fallen dramatically since Ben Katz's MIT work on mini cheetah. The actuators on the Unitree G1 based on that work are powerful for their size and near quasi-direct-drive."

Cautious optimism on the curve. FartyMcFarter took the comparison-to-LLMs view that got traction: "These robots look slow and not very fluid in their motions, but LLMs like ChatGPT also looked very dumb initially. If progress is as fast as LLMs, this could have massive applications in a few years."

The latency question. YuechenLi worried that a full LLM is the wrong control loop: "Running a full LLM to actuate a robot is way too heavy, the minimum latency they can get down to is probably 1-2 seconds even with powerful GPUs, which is not very useful for practical robotics applications." CardenB countered that specialized VLAs are already past that: "AFAIK NVidia has a VLA running on their chips with 2B params at 10Hz. They've openly published a .5B model running at 10Hz."

Marketing vs reality. bluber84 linked the developer blog and was blunt about the benchmark numbers: "Success-rate of ~60% Accuracy: ~80% That's pretty low and definitely not production ready." _davide_ was more cynical about access: "the model is real, but it feels 100% internal."

Hands-on signal. dr_blueberry reported early ER testing in a visual-agent harness: "Doing some initial testing of Gemini ER 2 within the Orion 2 visual agent harness... I'm impressed with how fast Gemini ER2 is."

aabhay asked for an honest read on "how much trouble do humanoids have with in the wild daily tasks like turning doorknobs, recovering from falls, avoiding knocking into things," and p1esk answered with the GPT framing: "It's still at GPT-1 level, but GPT-2 moment feels imminent." Meanwhile xnx noted that "while Anthropic and Open AI get 80% of the attention here," Google's span across frontier models, open weights, image, video, music, and robotics is quietly enormous.

Why It Matters for Developers#

For most developers, the immediate takeaway is that embodied reasoning is now a real API surface. Gemini Robotics ER 2 is available today in Google AI Studio and in private preview on the Gemini Enterprise Agent Platform, and the developer blog ships a robotics-overview doc, a getting-started notebook, and GitHub examples for configuring the model as a tool-orchestrating agent. You can build and test a "physical agent" loop - streaming video in, declaring VLA models or navigation APIs as tools, getting structured progress and decisions out - without owning a single robot. Robotics reasoning is becoming as accessible as any other Gemini API, while the VLA and on-device models stay behind DeepMind's early-access partner program.

The other thing worth watching is the benchmark honesty. DeepMind published per-task numbers, including the ugly ones (32% dustpan), and admits multi-finger dexterity is not solved. Compare that with the flattering aggregate success rates that used to ship with robotics demos, and it reads like a lab that expects to be judged on iteration speed rather than a single reveal. If the field follows the LLM trajectory, the 45% floor-pickup number will look either embarrassing or prescient in eighteen months.

Two caution notes from the thread are worth keeping. Flere-Imsaho flagged that the post never clarifies which models run locally versus in the cloud - relevant if the on-device tier is what makes home robots acceptable. And the safety framing matters: a robot that can clean a room can also knock someone over, and DeepMind's own safety work targets exactly that gap.

The wider trend is clear. Between Gemini Robotics 2, Mistral's navigation model, and the multimodal-and-robotics direction Black Forest Labs took with FLUX 3, the physical-AI lane is getting crowded, and the software is moving faster than the hardware.

Sources#

• Google DeepMind announcement: Gemini Robotics 2 brings whole body intelligence to robots
• Google developer blog: Introducing Gemini Robotics ER 2
• Hacker News discussion: https://news.ycombinator.com/item?id=49111237
• Safety technical report: Gemini Robotics 2: Safety Technical Report
• Robotics getting-started notebook: google-gemini/robotics-samples

Continue Reading#

• Mistral Releases Robostral Navigate: An 8B Robotics Navigation Model - a look at the other end of the physical-AI stack, where an 8B model handles robot navigation
• FLUX 3: Black Forest Labs Ships a Unified Multimodal Foundation Model for Image, Video, Audio, and Robotics - how a unified multimodal model is being aimed at robotics alongside media
• Gemma 4: The Open Model Guide for Developers - what Google DeepMind ships on the open-weights side of its strategy
• Gemini 3.5 Pro Developer Guide: 2M Context Window and Deep Think Mode - the current Gemini family ER 2 builds on
• Resource2Skill Turns Tutorials Into Agent Skills - how multimodal understanding is being packaged for agent use