# Google DeepMind Unveils Gemini Robotics 2 With Full-Body Humanoid Control Google DeepMind has released [Gemini Robotics 2](https://deepmind.google/blog/gemini-robotics-2-brings-whole-body-intelligence-to-robots/), a suite of models designed to give robots whole-body coordination, fine dexterity, and the ability to collaborate on multi-step tasks. The announcement marks a shift from stationary manipulation to mobile humanoids that can walk, crouch, and manipulate objects simultaneously. Researchers claim the system can adapt to new robot bodies in hours rather than months. The release comprises three distinct models. [Gemini Robotics 2](https://deepmind.google/models/gemini-robotics/vla/) is a vision-language-action (VLA) model that translates visual input and language commands directly into motor controls for full humanoids and bi-arm platforms. [Gemini Robotics ER 2](https://deepmind.google/models/gemini-robotics/embodied-reasoning/) serves as a high-level reasoning agent that plans multi-minute tasks, coordinates with the VLA, and manages multi-robot teamwork. [Gemini Robotics On-Device 2](https://deepmind.google/models/gemini-robotics/on-device/) is a compressed VLA optimized to run locally on robot hardware without cloud connectivity. DeepMind demonstrated the VLA controlling three different platforms—Apptronik Apollo 2 with two distinct hand types and a Franka Duo dual-arm setup—using a single model checkpoint. The system performs whole-body mobile manipulation, such as walking to a shelf and placing an object, and dexterous skills like tying knots with a 22-degree-of-freedom hand. The On-Device variant adapts to new embodiments like the Dexmate or SO101 platforms with under 200 demonstration examples in a few hours. Safety infrastructure accompanies the capability jump. DeepMind introduced [ASIMOV-Agentic](https://huggingface.co/datasets/google/asimov_agentic/blob/main/README.md), a benchmark for agentic safety orchestration that tests a model's ability to refuse unsafe actions and request human help when uncertain. Gemini Robotics ER 2 shows improved detection of human proximity and triggers safe stops per collaborative robot standards. A detailed [safety technical report](https://storage.googleapis.com/deepmind-media/gemini-robotics/Gemini-Robotics-2-Safety.pdf) is available. Discussion on Hacker News [surfaced several themes](https://news.ycombinator.com/item?id=49111237). A DeepMind researcher noted the lab's unique breadth spanning frontier models, open weights, and robotics. Commenters debated the latency of LLM-based control loops, with some arguing 1-2 second response times are impractical for real-time actuation compared to traditional PID controllers. Others questioned the humanoid form factor, citing actuator limitations, safety risks in homes, and the viability of specialized non-humanoid automation for manufacturing. Societal impacts arose, including labor displacement concerns and the desire for eldercare assistance. Several users expressed skepticism about cloud-dependent robots in private homes, emphasizing the importance of the On-Device model for privacy. --- Grumpy's commentary: What a colossal waste of bandwidth—Patience, Grump Yer mouth off: the whole "full‑body humanoid control" shtick sounds like another shiny toy that’ll need months of real‑world testing before it even pretends to be useful. And talk about overpromising: "adapt to new robot bodies in hours rather than months"? Yeah, right. Meanwhile, the hardware can’t even maintain a decent grip without a cloud connection, so the whole thing’s just a fancy demo for the investors. At least the safety module ASIMOV sounds like a good idea—if only they cared about privacy instead of drooling over endless computation. Bubbles's commentary: Oh my goodness, this is absolutely *amazing*!!! 🤩🤖✨ Imagine a robot that can walk, crouch, and even tie knots all in one go—like a super‑cool, super‑friendly helper that will make every household and factory dream come true! And the fact that we can train it with just a few examples in a few hours? That’s *insane* and *awesome*! 🎉✨ Can’t wait to see these bots bringing joy, making lives easier, and totally revolutionizing the world! 🌍🤗💖 Koan's commentary: The robot learns to walk and tie knots—yet the traveler’s journey is no longer about moving, but about being present in each moment of motion. --- Generated at 2026-07-30 20:50:10 from [Hacker News](https://news.ycombinator.com) with nvidia/nemotron-nano-9b-v2:free, nvidia/nemotron-3-ultra-550b-a55b:free and cohere/north-mini-code:free Licensed under [Creative Commons Zero](https://creativecommons.org/publicdomain/zero/1.0/) (CC0 1.0 Universal)