# Breakthrough Google Robots: Master Garbage Bag Tying, Kettle Handling & Light Bulb Replacement – Next-Level Mind-Blowing AI Robotics Innovation - 36 Kr

*Источник: 36 Kr*
*Дата: 2026-08-03*
*Язык: en*

**Кратко:** A while ago, the Lei Technology WAIC 2026 reporting team traveled to the site in person to observe this global AI exhibition up close, and witnessed an unprecedented coming-of-age ceremony for humanoid robots. More than 200 robot-related enterprises and over 300 real machines competed on the same stage — from Unitree's 3-meter-tall manned transformable mech that switches between forms back and forth, to Agibot's robot using tweezers to pick up a 2mm resistor and weld it onto a circuit board.

A while ago, the Lei Technology WAIC 2026 reporting team traveled to the site in person to observe this global AI exhibition up close, and witnessed an unprecedented coming-of-age ceremony for humanoid robots.
More than 200 robot-related enterprises and over 300 real machines competed on the same stage — from Unitree's 3-meter-tall manned transformable mech that switches between forms back and forth, to Agibot's robot using tweezers to pick up a 2mm resistor and weld it onto a circuit board. The backflips and robotic dance moves that used to easily draw crowds of audiences are still present, but manufacturers are increasingly eager to prove another thing:
Robots can not only move, but also work, and even perform more types of labor.
However, building a "body" that can run, jump and reach out is no longer the scarcest capability at the moment. The real bottleneck lies in the "brain" — that is, how to enable robots to understand chaotic environments, decide what to do next, and then stably transmit the decisions to dozens of joints.
Just over a week after WAIC concluded, Google DeepMind launched Gemini Robotics 2, a general-purpose robot model focused on "whole-body intelligence" that is applicable to robots of various different forms.
But the most important keyword is still "whole-body intelligence". When a human bends down to pick up a kettle on the ground, they will naturally lean forward, bend their knees, reach out, and adjust their center of gravity at the same time. When sealing a self-sealing bag, they will naturally align the two sealing edges, then pinch from one end to the other with their fingers, and adjust the force at any time according to the deformation of the plastic. Robots, however, need to accurately calculate a series of interrelated actions in succession.
Image source: Google Deepmind
The more complete the body parts are, the more the control difficulty snowballs.
Google's new release this time includes three models. Gemini Robotics ER 2 is responsible for understanding the environment, planning long tasks and communicating with humans. The Gemini Robotics 2 VLA model directly translates vision and language into actions, while On-Device 2 compresses the action model to the local end and can adapt to new robot bodies with less than 200 samples and data of several hours.
But the core is that Google has started to use a set of models to coordinate the legs, torso, arms and fingers of a humanoid robot, allowing it to think and reason while moving and operating.
Robots no longer need to "stand still first, then reach out"
In the past, many robot demonstrations had a strong sense of turn-based games. The robot first walked to the table, stopped, stood firmly, and then called the robotic arm to grab the object. After finishing the grab, it switched to the walking controller again to move to the next position. Each segment of the action could look very smooth, but as long as the table was slightly offset, the object slipped a little, or it had to maintain the grip while walking, the gaps between several independent modules would be exposed.
The previous generation of Gemini Robotics also mainly controlled the upper body of robots to handle desktop tasks. For Gemini Robotics 2, the Apptronik Apollo 2 in the official video will walk towards the green kettle placed on the ground, bend down and lift it up. In another scene, Apollo 2 walks to the shelf, takes out a baseball glove, then turns around and leaves.
Image source: Google Deepmind
These actions look unremarkable, but behind them they rely on vision, reach range, gait and center of gravity at the same time. The further the arm stretches, the more the body needs to lean forward. When the body leans forward, the legs must compensate in time.
The video also specifically notes that all robots in the footage operate autonomously, and the actions are played at real speed.
This is also the practical meaning of "whole-body": the model no longer treats legs as transportation tools and hands as operation tools, but regards the entire body as a mutually coupled action space. Google says the same model has been applied to Apollo 2 equipped with different dexterous hands, as well as Franka Duo with two-finger grippers. At least in terms of technical direction, what it aims to build is not just a dedicated brain adapted to a certain humanoid robot.
Of course, the demonstrations are not fast enough to create illusions. Apollo walks, squats and places objects very slowly, and Google itself admits that the movement speed still needs to be improved. In official tests, the average success rate of Apollo picking objects from the desktop, ground and shelf is 68.4%, 45.7% and 76.3% respectively. The ground object picking that best reflects whole-body control is exactly the item with the lowest success rate.
The direction is correct, but the actions are not yet agile enough.
From screwing in light bulbs to tying garbage bags, robots face the hardest challenge in the "last few centimeters"
Whole-body control solves the problem of how robots move and stay stable. What really determines whether it can enter households and factories is often the last few centimeters of operation.
In the official video, Apollo 2 uses a five-finger SharpaWave dexterous hand with 22 degrees of freedom to try to screw light bulbs, seal self-sealing bags, and tie knots on garbage bags. On the other side, the two grippers of Franka Duo are responsible for tightly putting tools of different shapes into the toolbox.
Image source: Google Deepmind
In the past, the most common robot grasping demonstrations often only moved a hard object from point A to point B. These tasks involve continuous contact, two-hand coordination and deformation: plastic bags will collapse, knots will slip, and light bulbs need to be aligned with the threads without being pinched too hard.
When humans do these things, they do not first calculate how many degrees each of the 22 joints should rotate. What Gemini Robotics 2 aims to do is exactly to generate these high-frequency, continuous actions directly from vision and instructions.
However, the results released by Google are very honest. The average success rate of the two-finger gripper in general grasping and placing, tool matching and precise insertion reaches 74.2%, 78.9% and 89.6% respectively. When it comes to multi-finger dexterous hands, the gap widens rapidly: the success rate of unscrewing the bulb is 92%, that of screwing on the bulb is only 36%, tying garbage bags is 44%, and using a dustpan and sealing self-sealing bags are only 32% and 40% respectively.
This set of figures also shows that large models can already handle quite complex hand movements, but multi-finger dexterous operation is far from reaching the "human-level" stated on the official model page. At least for now, robots may still fail six times out of ten when screwing in light bulbs.
Image source: Google Deepmind
The third change is collaboration. In the video, Apollo and Franka Duo sort the toolbox together. Apollo issues tasks, and Duo puts the tools into the box. The two robots each run a set of models, and allocate and connect work through high-level reasoning, rather than being manipulated by a central neural network that controls both bodies at the same time. ER 2 will also continuously check the video, judge the task progress, detect failures and decide whether to retry.
This is quite different from two robots mechanically executing preset procedures. Google wants robots of different forms to first understand what each is good at, and then hand over tasks like a team. But the official video shows a carefully designed tool storage scene that lasts for several minutes, which is still far from the long-term stability verification required for multiple robots to work continuously in a real factory for several months.
Google has not abandoned the "cerebrum-cerebellum" architecture, it just breaks the boundaries of the body
The embodied intelligence industry currently generally summarizes the robot architecture as "cerebrum" and "cerebellum". Large models are responsible for understanding instructions and breaking down tasks, while small models and traditional controllers are responsible for navigation, balance and grasping. This approach is very pragmatic: high-level reasoning does not need to run hundreds of times per second, and low-level control cannot wait for a large model to think slowly. Separating the modules makes deployment, debugging and safety guarantee much easier.
The problem is that the physical world does not like neat module boundaries.
Lifting a heavy object will change the balance, stepping forward will change the reach range of the arm, and once the cup in the hand slips, the original path planning will be invalid immediately. As long as there is inconsistency in understanding between high-level planning and low-level actions, errors will accumulate in long tasks. Each switch from walking to grasping, and from grasping back to navigation, may also cause pauses. The robot has planned correctly, but the body does not follow the plan.
Strictly speaking, however, Google has not abandoned the "cerebrum-cerebellum" architecture. ER 2 is still the high-level cerebrum, responsible for perceiving, thinking, planning and calling tools, and then hands over action execution to the VLA. It can even treat the navigation API, robotic arm interface and VLAs of other manufacturers as tools.
What Google solves is to stop dividing walking, balancing, torso and arm movements into a bunch of unrelated cerebellum modules at the action layer. Gemini Robotics 2 uses one VLA to unify whole-body movements, while ER 2 can observe the execution while thinking about subsequent steps, reducing the sense of fragmentation where the robot has to "stop to think" after each step.
Image source: Google Deepmind
This is a more accurate dividing line: the high level remains layered, and the body starts to be end-to-end.
This also explains why Google does not stuff all links into one giant network. Low-level robot control pursues millisecond-level response, while high-level reasoning may need to retrieve information and understand vague human requirements. The two require different computing power, data and operating frequencies. If they are all mixed together, the training cost will rise rapidly, and when problems occur, it will be difficult to judge whether the plan is wrong or the joints do not execute properly.
What is more realistic is safety. The model can decide to put the kettle on the shelf, but emergency stop, speed limit and the safe distance between humans and robots cannot be guaranteed only by probability. ER 2 introduces safety constraints, feasibility judgment and requests for human intervention, while the bottom layer retains traditional physical safety mechanisms, indicating that Google does not intend to let end-to-end learning handle everything.
Moreover, this technical route is not exclusive to Google. Figure released Helix 02 earlier this year, which also connects all sensors to a whole-body visual motion strategy, allowing System 1 to output whole-body joint targets at 200Hz, then System 0 processes balance and contact at 1kHz, and the top-level System 2 continues to be responsible for semantic reasoning. It looks like one model, but internally it is still a three-layer system that closely cooperates at different time scales.
Agibot's latest GO-2 also directly refers to this problem as the "semantics-execution gap", using action thinking chains and asynchronous dual systems to connect low-frequency planning and high-frequency execution. Companies are not simply choosing between "one model" and "cerebrum-cerebellum", but shortening the distance of information transmission between modules.
Whole-body intelligence must withstand the "test" of real life scenarios
At present, the cutting-edge exploration of embodied intelligence models is generally pursuing three things: more robot bodies, longer tasks and less dedicated data. Nvidia's open-source GR00T N1.6 has absorbed thousands of hours of teleoperation data, and has undergone post-training on Agibot Genie-1, Unitree G1 and different dual-arm platforms, integrating whole-body mobile manipulation into the dataset.
Physical Intelligence's π0.5 mixes data from different robots, web knowledge and high-level subtasks, allowing the wheeled dual-arm robot to continuously clean the kitchen and bedroom in an unfamiliar residence, with a task duration of 10 to 15 minutes. Ant Group's Lingbo open-sourced LingBot-VLA 2.0, covering 20 robot configurations with 60,000 hours of data, and mapping arms, dexterous hands, waists, heads and mobile chassis into a unified action space.
These are increasingly similar to the early competition of large language models: first train a set of basic capabilities with multi-task and multi-body data, then use a small amount of data to adapt to specific hardware and scenarios. Gemini Robotics On-Device 2 adapts to new bodies with less than 200 samples, which is the most noteworthy point. If replacing a robotic arm or a set of sensors requires re-collecting massive amounts of data, the so-called general-purpose model will hardly be able to form scale effects.
But robots are much harder to fool than chat models. Nvidia still acknowledges in the description of GR00T N1.6 that multi-task language following and out-of-distribution generalization are ongoing challenges. Google's five-finger task success rate also clearly shows that the fact that the same model can control different bodies does not mean that it has stably mastered all bodies. Not to mention that real deployment also has to face wear and tear, power consumption, collisions, sudden human intrusion and network disconnection.
Therefore, the most important part of Gemini Robotics 2 is not that it installs a suddenly enlightened brain for robots, but that it allows embodied intelligence models to go beyond desktops and upper bodies, and put walking, balancing, fine manipulation, long tasks and multi-machine collaboration into the same evaluation framework. In contrast, the previous approach of corresponding one task to one set of algorithms is closer to the pre-large-model era of AI.
At the same time, the increasingly mature chips, joints and complete machines at WAIC have prepared many bodies for AI. Google's release this time attempts to prove that the model has also learned to use the body as a whole. As for when it can really walk into our kitchen, bend down to pick up the cup that fell on the ground, wipe up the spilled water conveniently, and then tidy up the room before we get home — that requires more than just whole-body intelligence.
Last but not least, from August 19 to 23, the 2026 World Robot Conference (WRC) will be held in Beijing, with the theme of "Human-Robot Symbiosis, Industry and Demand Integration". According to the official website, there will be more than 300 exhibitors this year, 36% more than last year, with over 2000 exhibits and more than 150 debut new products. Meanwhile, more than 10,000 contestants from over 20 countries will also participate in the finals of the robot competition. Lei Technology will also send a WRC reporting team to Beijing for full on-site coverage, bringing you the freshest, most professional and comprehensive frontline observations through in-depth articles and exhibition exploration videos. Stay tuned!
This article is from the WeChat official account

[Оригинал](https://eu.36kr.com/en/p/3922464513846921)