Keerthana Gopalakrishnan, research lead for Gemini Robotics at Google DeepMind, had not been following China's humanoid races until her father brought them up. "Wow, did you see those humanoids running?" he asked. Then the videos found her: humanoids running, and some catching fire. Friends in India who have gone into politics and student protests told her that China was building robots that could run faster than the fastest humans.
Gopalakrishnan spoke on The Cognitive Revolution in an episode published October 3, 2026. She said that from a research perspective the result was "not very surprising." In her account, running fast on a flat track is the kind of skill simulation teaches well. Manipulation, the work of grasping and handling objects, is not, and that harder work is where her team spends its time.
The race and the collage
The event was the second World Humanoid Robot Games. Xinhua reported that Team Tianjiao's Tiangong Ultra robots won gold and silver in the large-robot 100-meter final. The winning time was 8.64 seconds, a games record. Xinhua also described robots collapsing after the finish, with some carried away on stretchers.
Gopalakrishnan said the competition "really captured the public imagination in a very emphatic way." She also said running works as a benchmark: a way to see where humans are and where robots are. She pointed to collages people had made comparing the races with the DARPA Robotics Challenge. As she remembered it, those robots "would just walk, open a door and then like really fall down." She put that challenge about 20 years back. DARPA's announcement dates its finals to June 2015, about eleven years before the games. That contest was about disaster response. Qualification tasks included walking ten meters without falling, getting up from a prone position, crossing a barrier and turning a valve a full 360 degrees. "So it has been amazing progress," she said.
Why speed is the easy part
The conversation then turned to whether a humanoid that runs 20 miles an hour is useful at all. Gopalakrishnan called the sprint a display of the state of the art in bipedal, or two-legged, locomotion. She allowed that speed might matter "for some defense applications." But her team, she said, is focused on robots that help people and do useful things in the physical world. "I think I'm a very productive human," she said. "A lot of my friends are also very productively employed, but we don't run faster than Usain Bolt."
She said locomotion has advanced so far partly because of sim-to-real transfer. That is the question of whether a skill a robot learns in a simulated world still works on a real machine. Many of these locomotion models, she said, are trained in simulation. The discussion noted one reason why: teams could not keep training by crashing real robots into walls, because they would run out of robots.
The physics explains which skills transfer well. For walking and running, Gopalakrishnan said, the ground or a wall can be modeled as a surface the robot touches, and usually a flat one. That makes locomotion a very good task for simulation. Manipulation is different. "What you care about is contact," she said, "and when contact happens and how objects behave, that's where current simulations start to break a little bit."
She ranked tasks by how well they simulate. Cloth folding is harder because cloth has friction. Much simple pick-and-place, by contrast, can be solved in simulation, because the contact dynamics are fairly predictable. When the discussion summed this up as the more rigid the body, whether floor or wooden cube, the easier the simulation, she agreed: contact physics is hard to model, she said, so "where the physics is easier to model, simulation gets it."
Gopalakrishnan was careful when asked how the racing robots were actually controlled. "I didn't work on these demos, so I'm only speculating," she said. She pointed to a large published literature on training locomotion in simulation. Her guess was that the robots used reinforcement learning (RL) controllers, software that learns movements by trial and error toward a reward. Such controllers would probably first learn from motion imitation, then be fine-tuned with RL to fit the robot's body and the single goal of running fast. There might be some navigation too, she said, "although I would think that just running a race is quite easy."
A frontier that keeps moving
Asked whether she hands off any everyday chores to robots, Gopalakrishnan said that as a researcher she works on whatever robots cannot yet do. "The frontier is moving," she said. When she started working on manipulation, even pick-and-place was very hard, and the right algorithms did not exist. With foundation models, the large general-purpose AI models trained on broad data, that kind of task is now "quite easy." Her team's demos used to involve gripper robots, which have simple pincer-like ends. Now they involve humanoids with hands.
Her example of the new frontier was Gemini Robotics 2, which Google announced on July 30, 2026. She described it as whole-body manipulation with generalization. The robot takes steps and squats while handling objects, adjusts continuously to what it senses ("in a closed loop sort of way") and reasons about different kinds of objects. A year earlier, she said, she would have been "a bit surprised" to hear this would be possible by now. Her team began working on humanoids about two years ago, when "trying to just grab an apple" was really hard. Now, she said, you can prompt it. She mentioned a video she posted on Twitter of herself playing with the robot as it squatted in different places to pick up a watering can. Google's announcement describes a similar watering-can task: a humanoid walks to a table, picks up the can, carries it to shelving and puts it in a green bin on the bottom shelf.
Google's announcement also gives a sense of how far this has come and how far it has to go. Its results charts show average success rates over multiple tasks within each skill category, and Google says Gemini Robotics 2 reaches a medium to high success rate on whole-body and gripper-based dexterous tasks, while multi-finger dexterous manipulation remains challenging.
From grippers to fingers
The conversation then turned to robot hands. Gopalakrishnan traced the change through her team's two releases. The first Gemini Robotics, announced in March 2025, showed dexterity with grippers, such as folding origami and packing food into a ziplock bag. Gemini Robotics 2 showed multi-fingered dexterity, which she said "can tie trash bags and stuff" and control multi-fingered hands very finely. She called the gap "a year and a quarter"; the two announcements were about 16 months apart.
"Here's where my internal model is also updating," she said. Hand robots can now do everything gripper robots could do when grippers were at the cutting edge of dexterity, she said, plus things grippers could not. "Hands are now at the frontier of dexterity and grippers are maybe not." She said she was "quite surprised" by how far hand hardware has come, and that many good hands on the market are now used in research. But they still need work to become very reliable and very repeatable, and to get cheaper.
Google's announcement shows the same unevenness. It says the model can control the five-fingered, 22-degree-of-freedom SharpaWave hand on the Apollo 2 robot to tie knots or seal a ziplock bag, yet it also says multi-finger dexterous manipulation remains challenging.
Can it open a jar?
Asked whether a robot hand could stand in for a person on jar duty, Gopalakrishnan said it depends on the hand, and that off-the-shelf hands vary a lot. She said the Wuji hand is closer to "maybe like a 10-year-old," a bit weaker and closer to human size. Wuji's own specifications list 20 actively controlled degrees of freedom and a fingertip pressing force of 15 newtons. They also list a 10-kilogram static load for the whole hand, a separate measurement from fingertip force.
She said a second, stronger hand could lift about 20 kilograms, and she did not offer that as a measured figure. The transcript renders its name as "sharper"; the likely candidate is Sharpa's hand, which Google used in its Apollo tests. "I don't know how much the torque stuff, but I've seen it open jars," she said, adding, "Maybe not a very tight one." She said it is much bigger than her own hand. Sharpa, for its part, described its SharpaWave in a December 2025 announcement as a human-sized hand with 22 active degrees of freedom that had reached mass production. Some hands, Gopalakrishnan said, are simply built to lift more. The conversation joked that the last job left for humans might be opening jars too stuck for robots.
Soft hands, gloves and touch
Asked about soft robotics, Gopalakrishnan said it covers a wide range, from soft robot bodies to very soft hands. What matters for AI-driven robotics, she said, is what is repeatable and durable. The trend she finds most interesting is "the UMI type of effort": gloves and other tools for collecting demonstrations, and touch sensing as an input. The Universal Manipulation Interface (UMI), developed by Chi and colleagues, lets people collect demonstrations with handheld grippers instead of a robot. The grippers have flexible fingers and wrist-mounted cameras, and the way the fingers bend gives indirect feedback about contact. Touch sensing is moving into commercial hands too. Sharpa says its fingertips combine a miniature camera with more than a thousand tactile pixels per finger, which it presents as a way to adjust grip and detect slipping. Gopalakrishnan said that building such systems well is still an area of research.
She added that she is not fully versed in soft robotics, and that roboticists come in different kinds. Some very mechanical-engineering-focused roboticists look at her work and tell her she is not a real roboticist. "And I'm like, sure, okay," she said. She named ICRA, the International Conference on Robotics and Automation, as a place where "there's every kind of roboticist."