# Google DeepMind Launches Gemini Robotics 2 for Full-Body Humanoid Control > Source: > Published: 2026-07-31 12:53:58.403409+00:00 # Google DeepMind Launches Gemini Robotics 2 for Full-Body Humanoid Control - Google DeepMind shipped three physical AI models — Gemini Robotics 2, Gemini Robotics ER 2, and Gemini Robotics On-Device 2 — on July 30. [[1]](https://blog.google/innovation-and-ai/models-and-research/google-deepmind/gemini-robotics-er-2/) - Gemini Robotics 2 is the first VLA model to control a full humanoid end-to-end, demonstrated on Apptronik's Apollo 2 with tasks like walking, retrieving objects, and shelf placement. [[3]](https://theaiinsider.tech/2026/07/30/google-introduces-gemini-robotics-2-with-whole-body-intelligence/) - ER 2 enables multi-robot collaboration and handles task sequences of hundreds of steps, with the ability to resume from the last successful step after errors. [[2]](https://siliconangle.com/2026/07/30/google-deepmind-debuts-gemini-robotics-2-model-series-humanoid-robots/) - Benchmark results include 92% success on light-bulb unscrewing and 89.6% on precise insertion tasks with a two-finger gripper. [[3]](https://theaiinsider.tech/2026/07/30/google-introduces-gemini-robotics-2-with-whole-body-intelligence/) - Alphabet shares rose approximately 3% in pre-market trading on July 31 following the announcement. [[5]](https://www.tradingkey.com/analysis/stocks/us-stocks/262067418-stock-google-alphabet-goog-gemini-llm-price-predittion-ai-tradingkey) Google DeepMind on July 30 released a family of three AI models designed to give robots full-body coordination, multi-step reasoning, and the ability to work alongside other machines. The suite — Gemini Robotics 2, Gemini Robotics ER 2, and Gemini Robotics On-Device 2 — marks the lab's most ambitious push into physical AI, extending control from tabletop manipulation to whole-body humanoid locomotion. [[1]](https://blog.google/innovation-and-ai/models-and-research/google-deepmind/gemini-robotics-er-2/) The flagship Gemini Robotics 2 is a vision-language-action model that converts visual and language inputs directly into motor commands across an entire humanoid body. In demonstrations on Apptronik's Apollo 2 robot, the system walked to a table, retrieved a watering can, navigated to a shelf, and placed it precisely — coordinating legs, torso, arms, and 22-degree-of-freedom SharpaWave hands under a single learned policy. [[3]](https://theaiinsider.tech/2026/07/30/google-introduces-gemini-robotics-2-with-whole-body-intelligence/) Gemini Robotics ER 2, the embodied reasoning layer, acts as a high-level planner that distributes work across multiple robots, tracks progress through camera feeds, and can call external cloud services including Google Search for real-time information. Built on the Gemini 3.5 Flash architecture with a 128,000-token context window, it is available now through Google AI Studio, the Gemini API, and the Gemini Enterprise Agent Platform in private preview. [[2]](https://siliconangle.com/2026/07/30/google-deepmind-debuts-gemini-robotics-2-model-series-humanoid-robots/)[[4]](https://deepmind.google/models/model-cards/gemini-robotics-er-2/) ## The Three Models The release comprises three distinct models, each targeting a different layer of the robotics stack. Gemini Robotics 2, the vision-language-action model, handles direct motor control and dexterous manipulation. Gemini Robotics ER 2 sits above it as the reasoning and orchestration layer, managing task planning, human communication, and multi-robot coordination. Gemini Robotics On-Device 2 runs locally on robot hardware, reducing dependence on cloud connectivity and requiring only hours of training to adapt to new platforms. [[2]](https://siliconangle.com/2026/07/30/google-deepmind-debuts-gemini-robotics-2-model-series-humanoid-robots/)[[3]](https://theaiinsider.tech/2026/07/30/google-introduces-gemini-robotics-2-with-whole-body-intelligence/) The VLA model has been tested across multiple robot embodiments including Apptronik's Apollo 2 humanoid with both SharpaWave and Inspire hands, the Franka Duo with a Robotiq gripper, and platforms from Dexmate, SO101, and Trossen. The on-device model can adapt to new two-arm robot designs with fewer than 200 training examples. [[3]](https://theaiinsider.tech/2026/07/30/google-introduces-gemini-robotics-2-with-whole-body-intelligence/) ## Benchmark Results Google DeepMind published detailed performance metrics across several task categories. On the Apollo 2 humanoid, the system achieved a 76.3% success rate on shelf pickups, 68.4% on table pickups, and 45.7% on floor pickups — tasks that demand whole-body coordination including crouching and balance management. [[3]](https://theaiinsider.tech/2026/07/30/google-introduces-gemini-robotics-2-with-whole-body-intelligence/) Dexterous manipulation proved more uneven. The system hit 92% on unscrewing a light bulb but only 36% on screwing one in. Trash bag tying came in at 44%, bag sealing at 40%, and dustpan use at 32%. On the Franka Duo platform with a two-finger gripper, results were stronger: 89.6% on precise insertion, 78.9% on tool kitting, and 74.2% on general pick-and-place. [[3]](https://theaiinsider.tech/2026/07/30/google-introduces-gemini-robotics-2-with-whole-body-intelligence/) ## Partners and Availability Google DeepMind is making Gemini Robotics 2 and On-Device 2 available to more than 100 trusted testers and early partners. Hardware partners include Apptronik, which opened a shared 90,000-square-foot 'Robot Park' training facility with Google DeepMind in June 2026, as well as Franka and Agile Robots. Boston Dynamics is also acknowledged in the technical credits, consistent with an existing agreement announced in May 2026 that gave DeepMind access to a fleet of Atlas robots. [[3]](https://theaiinsider.tech/2026/07/30/google-introduces-gemini-robotics-2-with-whole-body-intelligence/)[[6]](https://roboticsandautomationnews.com/2026/07/06/apptronik-launches-robot-park-to-train-apollo-humanoid-robots-with-google-deepmind/103069/) ER 2 has the broadest availability of the three, accessible through Google AI Studio, the Gemini API, and the Gemini Enterprise Agent Platform in private preview. The VLA model is limited to early-access partners through a sign-up process. [[2]](https://siliconangle.com/2026/07/30/google-deepmind-debuts-gemini-robotics-2-model-series-humanoid-robots/)[[4]](https://deepmind.google/models/model-cards/gemini-robotics-er-2/) ## Safety and Competitive Context Alongside the models, Google DeepMind introduced ASIMOV-Agentic, a benchmark for evaluating embodied reasoning safety. The framework tests a robot's ability to reject unsafe actions, recognize uncertainty, and request human assistance. The model card for ER 2 explicitly restricts use in safety-critical domains including healthcare and transportation where failures could cause injury. [[2]](https://siliconangle.com/2026/07/30/google-deepmind-debuts-gemini-robotics-2-model-series-humanoid-robots/)[[4]](https://deepmind.google/models/model-cards/gemini-robotics-er-2/) The release intensifies competition in the fast-moving physical AI sector. Nvidia, Tesla, and a growing roster of startups including Figure and 1X are pursuing similar goals of general-purpose humanoid robot intelligence. Google's approach differs by providing the AI software layer rather than building robot hardware, positioning it as a platform provider across multiple robot embodiments. [[3]](https://theaiinsider.tech/2026/07/30/google-introduces-gemini-robotics-2-with-whole-body-intelligence/) ## Market Reaction Alphabet shares rose approximately 3% in pre-market trading on July 31, briefly reaching $338.95. The stock had been trading near $315 as recently as July 24 amid broader concerns about big-tech AI capital expenditure efficiency. Analysts maintain a consensus Strong Buy rating on GOOGL with an average price target of $427.81, according to TipRanks data. [[5]](https://www.tradingkey.com/analysis/stocks/us-stocks/262067418-stock-google-alphabet-goog-gemini-llm-price-predittion-ai-tradingkey) ## Companies mentioned ## Further sources [[1] Google DeepMind blog — Gemini Robotics ER 2: powering robotics with video under… ↗](https://blog.google/innovation-and-ai/models-and-research/google-deepmind/gemini-robotics-er-2/) [[2] SiliconANGLE — Google DeepMind debuts Gemini Robotics 2 model series for humano… ↗](https://siliconangle.com/2026/07/30/google-deepmind-debuts-gemini-robotics-2-model-series-humanoid-robots/) [[3] The AI Insider — Google Introduces Gemini Robotics 2 with Whole Body Intelligen… ↗](https://theaiinsider.tech/2026/07/30/google-introduces-gemini-robotics-2-with-whole-body-intelligence/) [[4] Google DeepMind — Gemini Robotics ER 2 Model Card ↗](https://deepmind.google/models/model-cards/gemini-robotics-er-2/) [[5] TradingKey — Alphabet Stock Forecast: New AI Robot Model Fuels 3% Rally ↗](https://www.tradingkey.com/analysis/stocks/us-stocks/262067418-stock-google-alphabet-goog-gemini-llm-price-predittion-ai-tradingkey) [[6] Robotics and Automation News — Apptronik launches Robot Park to train Apollo hu… ↗](https://roboticsandautomationnews.com/2026/07/06/apptronik-launches-robot-park-to-train-apollo-humanoid-robots-with-google-deepmind/103069/) The stories that matter, in one email. Free — unsubscribe anytime.