At WAIC 2026, more than 200 robotics companies and over 300 physical robots competed on the same stage, showcasing capabilities ranging from backflips to precision welding. But the industry’s real bottleneck is no longer the “body”\u2014it’s the “brain.” Just over a week after the expo ended, Google DeepMind released Gemini Robotics 2, a general-purpose robot model centered on “whole-body intelligence,” aiming to teach robots to coordinate their entire bodies like humans and conquer the “last few centimeters” of fine manipulation.<\/p>\n
Gemini Robotics 2 comprises three models: Gemini Robotics ER 2 handles environment understanding, long-horizon task planning, and human communication; Gemini Robotics 2 is a vision-language-action (VLA) model that translates visual and language inputs directly into actions; and On-Device 2 compresses the action model for local deployment, adapting to new robot platforms with fewer than 200 examples and just hours of data. The core idea: Google is using a unified set of models to coordinate a humanoid robot’s legs, torso, arms, and fingers, enabling it to think, reason, and move simultaneously.<\/p>\n
In the past, robot demos often had a “turn-based game” feel: the robot walks to the table, stops, stabilizes, then calls a robotic arm routine to grasp an object, then switches to a walking controller to move to the next location. This modular architecture exposes its seams when the table shifts, objects slide, or the robot needs to grasp while walking. The previous generation, Gemini Robotics, primarily controlled the upper body for tabletop tasks. In Gemini Robotics 2’s official video, an Apptronik Apollo 2 walks toward a green water bottle on the floor, bends down to pick it up, then walks to a shelf and retrieves a baseball glove\u2014the motions fluid and continuous, with the video explicitly labeled as autonomous operation at real speed.<\/p>\n
Google says the same model has been deployed on the Apollo 2 with different dexterous hands, as well as on the Franka Duo with two-finger grippers, suggesting it’s not a dedicated brain for one specific humanoid but an attempt at a general-purpose “whole-body intelligence” solution. That said, Apollo’s movements in the demo remain slow, and Google acknowledges that locomotion speed needs improvement. In official tests, Apollo’s average success rates for grasping from tabletops, the floor, and shelves were 68.4%, 45.7%, and 76.3%, respectively\u2014with floor grasping, the task most dependent on whole-body control, posting the lowest success rate.<\/p>\n
Whole-body control solves how a robot gets somewhere and stays balanced, but what truly determines whether it can enter homes and factories is often the “last few centimeters” of fine manipulation. In the official video, the Apollo 2 uses a SharpaWave five-finger dexterous hand with 22 degrees of freedom to screw in a light bulb, seal a zip-lock bag, and tie off a trash bag; the Franka Duo’s two grippers pack oddly shaped tools tightly into a toolbox. These tasks involve continuous contact, bimanual coordination, and deformation\u2014plastic bags collapse, knots slip, and light bulbs require precise thread alignment without excessive grip force. Gemini Robotics 2’s goal is to generate high-frequency continuous actions directly from vision and instructions, rather than precomputing every joint angle.<\/p>\n
Google’s published success rates are refreshingly candid: the two-finger gripper averaged 74.2% for general pick-and-place, 78.9% for tool matching, and 89.6% for precision insertion. But the dexterous hand’s performance varies widely: unscrewing a light bulb hit 92%, while screwing one in was just 36%; tying a trash bag reached 44%, and using a dustpan and sealing a zip-lock bag managed only 32% and 40%, respectively. These numbers show that large models can now take over complex hand movements, but multi-finger dexterous manipulation is far from “human-level”\u2014a robot screwing in a light bulb may still fail six times out of ten.<\/p>\n
Another highlight of Gemini Robotics 2 is collaboration. In the video, Apollo and Franka Duo work together to organize a toolbox: Apollo issues the task, Duo packs tools into the box, and both robots run the same model, coordinating through high-level reasoning rather than a central neural network controlling both bodies simultaneously. ER 2 continuously reviews video to assess task progress, detect failures, and decide whether to retry. This differs from mechanically executing preset routines\u2014Google wants robots of different form factors to understand each other’s strengths and hand off tasks like a team. However, the official video shows a designed, few-minute tool-organizing scenario; validating continuous multi-robot collaboration in real factories over months remains a long road.<\/p>\n
The embodied AI industry commonly divides robot architecture into a “brain” and a “cerebellum”: large models handle instruction understanding and task decomposition, while smaller models and traditional controllers manage navigation, balance, and grasping. This modular design is pragmatic and easy to deploy, but the physical world doesn’t always respect module boundaries\u2014lifting a heavy object changes balance, taking a step changes the arm’s reachable range, and a slipping cup invalidates path planning. If high-level plans and low-level actions don’t align, errors compound over long tasks. Google hasn’t abandoned the “big brain\/small brain” hierarchy\u2014ER 2 remains the high-level brain, responsible for seeing, thinking, planning, and calling tools\u2014but Gemini Robotics 2 uses a single VLA to unify whole-body actions, while ER 2 observes execution and plans next steps, reducing the jarring “stop and think” disconnect. It’s a cleaner dividing line: the high level stays layered, but the body goes end-to-end.<\/p>\n
Google also isn’t cramming everything into one giant network, because low-level control requires millisecond-level response, while high-level reasoning may involve retrieving information and understanding ambiguous instructions\u2014each with different demands on compute, data, and frequency. ER 2 introduces safety constraints, feasibility checks, and requests for human intervention, while low-level traditional physical safety mechanisms remain in place, indicating Google isn’t betting on end-to-end learning to handle everything.<\/p>\n
This approach isn’t unique to Google. Figure’s Helix 02, released earlier this year, feeds all sensors into a whole-body visuomotor policy: System 1 outputs whole-body joint targets at 200Hz, System 0 handles balance and contact at 1kHz, and System 2 handles semantic reasoning\u2014a three-tier system operating on different timescales. AgiBot’s GO-2 addresses the “semantic-execution gap” with action chain-of-thought and asynchronous dual systems connecting low-frequency planning to high-frequency execution. The industry isn’t simply choosing between “one model” and “big brain\/small brain”\u2014it’s shortening the distance of communication between modules.<\/p>\n
Current frontier research in embodied AI models focuses on three things: more embodiments, longer tasks, and less task-specific data. Nvidia’s open-source GR00T N1.6 incorporates thousands of hours of teleoperation data, with post-training on platforms like AgiBot’s Genie-1 and Unitree’s G1, and includes whole-body mobile manipulation in its dataset. Physical Intelligence’s \u03c00.5 mixes data from different robots, web knowledge, and high-level subtasks, enabling a wheeled dual-arm robot to continuously clean a kitchen and bedroom in an unfamiliar home for 10 to 15 minutes per task. Ant Lingbo’s open-source LingBot-VLA 2.0 uses 60,000 hours of data covering 20 robot configurations, mapping arms, dexterous hands, waists, heads, and mobile bases into a unified action space.<\/p>\n
These developments increasingly resemble the early days of large language model competition: first train foundational capabilities on multi-task, multi-embodiment data, then adapt to specific hardware with minimal data. Gemini Robotics On-Device 2’s ability to adapt to a new platform with fewer than 200 examples is the key\u2014if swapping a robotic arm or sensor requires recollecting massive datasets, general-purpose models can’t achieve scale effects. But robots are harder to “fool” than chatbots. Nvidia still acknowledges in GR00T N1.6 that multi-task language following and out-of-distribution generalization remain ongoing challenges, and Google’s five-finger success rates show that a model controlling different bodies doesn’t mean it has reliably mastered all of them. Real-world deployment must also contend with wear and tear, battery life, collisions, human intrusion, and network outages.<\/p>\n
The most significant takeaway from Gemini Robotics 2 isn’t that it gives robots a suddenly enlightened brain\u2014it’s that embodied AI models are beginning to move beyond tabletops and upper bodies, putting walking, balance, fine manipulation, long-horizon tasks, and multi-robot collaboration on the same exam. By contrast, the previous “one task, one algorithm set” approach was closer to the pre-LLM era. The increasingly mature chips, joints, and full robots at WAIC have prepared multiple “bodies” for AI; Google is trying to prove that models can also learn to use the body as a whole. As for when robots will walk into kitchens, bend down to pick up a cup, and wipe away the water stains\u2014that requires not just whole-body intelligence, but the maturation and validation of many more technologies.<\/p>\n","protected":false},"excerpt":{"rendered":"At WAIC 2026, more than 200 robotics companies and over 300 physical robots competed on the same stage,…\n","protected":false},"author":2,"featured_media":127489,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[9],"tags":[35207,63828,62787,5044,63827,63825,62763,132,7543,58,6019,63826],"class_list":["post-127488","post","type-post","status-publish","format-standard","has-post-thumbnail","category-google","tag-agibot","tag-ant-lingbo","tag-apptronik-apollo-2","tag-deepmind","tag-figure","tag-franka-duo","tag-gemini-robotics-2","tag-google","tag-google-deepmind","tag-nvidia","tag-physical-intelligence","tag-sharpawave"],"_links":{"self":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts\/127488","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/comments?post=127488"}],"version-history":[{"count":0,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts\/127488\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/media\/127489"}],"wp:attachment":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/media?parent=127488"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/categories?post=127488"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/tags?post=127488"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}