Gemini Robotics 2.0: DeepMind's New Robot Brain

This summer, a robotics event in China went viral for a simple reason: humanoid robots ran faster than Usain Bolt. The clips spread everywhere, carrying one message: the machines have arrived. But according to Keerthana Gopalakrishnan, staff research scientist at Google DeepMind and research lead for Gemini Robotics, that spectacle is close to beside the point. "I'm a very productive human. A lot of my friends are also very productively employed, but we don't run faster than Usain Bolt," she says (07:18). Her point is that foot speed almost never limits what a robot can do for a person. What limits it is manipulation, the far harder job of making hands and fingers grip, twist, and place real objects without crushing or dropping them.

The reason running went viral first is also the reason it was never going to be the real test. Locomotion is, in Gopalakrishnan's words, easy to train in simulation, because the ground is mostly flat and rigid, and contact with it is predictable (08:55). Manipulation is a different animal. The moment an object deforms, like cloth folding over itself, friction and contact physics get messy, and today's simulators start to break (09:23). Pick and place tasks with rigid objects, cubes and cans, still simulate fine. Cloth, and anything springy or crumpled, does not.

That gap between what simulates well and what actually matters in a kitchen or a warehouse is the quiet theme of the whole conversation. It is also why Gopalakrishnan, despite being visibly proud of her team's recent progress, still rates the field as being in its "GPT-2 era" (20:13). She does not mean the models are unsophisticated. She means two specific things are still missing: reliable few shot learning across a wide range of tasks, and what she calls cross-embodiment, the ability of one brain to control very different robot bodies without starting over (20:41). A laptop runs the same software whether it is a Mac or a PC. A robot foundation model, by contrast, can still be helpless the moment someone swaps in an unfamiliar robot. Until that stops being true, she argues, it is hard to call the underlying model a generic brain.

The Architecture Behind the Demo

This year DeepMind released Gemini Robotics 2.0 to answer exactly that question, and the shape of the system maps neatly onto the old split between thinking and doing. The suite has three parts. Gemini Robotics ER2, for embodied reasoning, is the slow, deliberate layer, built on the Gemini Flash line, and it handles reading instruments, pointing at objects, and parsing what a person actually wants (24:40). It is available through DeepMind's API, so outside developers can wire it up to their own robots. Below that sits Gemini Robotics 2, the vision-language-action model, which turns ER2's instructions into actual joint movements. It now controls the whole humanoid body from fingertips to feet in a single closed loop, rather than handing stabilization to a separate, simpler controller (50:05). A third, smaller version, Gemini Robotics On-Device, compresses the same capability so it can run on the robot's own local computer, with no cloud connection required (25:36). That matters anywhere network latency is a problem, since a round trip to a cloud model, which the host's own tests sometimes clocked at six seconds, is simply too slow for some tasks (26:41).

The ER2 model's working memory has a concrete limit worth knowing. Its context window, the amount of past information it can hold while reasoning, is 128,000 tokens, which Gopalakrishnan translates to roughly three minutes of densely packed robot memory (32:42, 33:05). Past that, the system has to compress. Her advice for anyone building on top of it is not to simply keep more video frames but to summarize in text instead, the way a person recalls flipping an egg without replaying every frame of the pan (34:22, 35:08). Text is more compressible than images, and it still carries the narrative of what happened.

"Working on generalization is like lifting the boat, lifting the wave for all the boats." (Keerthana Gopalakrishnan, 15:17)

The instruction channel between ER2 and the VLA is still fairly narrow, mostly language plus pointing at objects in an image (46:59). That is a deliberate research stage, not a permanent ceiling. DeepMind calls this problem steerability, and the team is actively trying to widen that bandwidth, because right now, if the connection between the reasoning model and the action model is only text, some nuance of a scene can get lost in translation (47:08).

Hands, Safety, and the Next Frontier

Perhaps the clearest sign of progress is hardware. A year ago the frontier was grippers, simple two-fingered clamps. Gemini Robotics 1, released in spring of last year, showed real dexterity, but with grippers. By this summer, Gemini Robotics 2 was tying trash bags and performing fine multi-fingered control (57:45). Gopalakrishnan now says hands, not grippers, are where the dexterity research happens, since hand-equipped robots can do everything grippers could plus tasks grippers cannot attempt at all (58:14). Capability still varies a lot by hardware: the Sharper hand can lift about 20 kilograms and has opened jars, while the Wuji hand is closer in strength to a ten year old's grip (59:08, 59:31).

Underneath this sits a less flashy but more consequential finding. With around 200 examples, a general foundation model can learn a new task on a new robot body, because it already carries broad physical common sense and multi-embodiment control data to build on (51:47). That is the clearest evidence of cross-embodiment working in practice, though Gopalakrishnan adds a caveat: taking a brand new robot body and getting high reliability on many tasks with zero extra data still has no strong precedent (51:18). One brain, instantly fluent in any body, is closer than it was, but not yet real.

Safety, in this telling, is not a brake on progress but a feature competing for the same engineering attention as everything else. "I think of safety as, like, a capability, right?" she says (63:32), arguing that people simply will not adopt robots that are unsafe, so the most useful systems will also turn out to be the safest ones. She separates this from the familiar AI safety conversation about goals and guardrails. Robotics has its own category, operational safety, making sure a robot does not hurt someone through clumsiness rather than malice (64:27). In one internal demo, someone dropped a basket over a working humanoid's head mid-task. The right response was not to blunder on blind but to notice the obstruction and ask for help removing it (65:08).

That example points to something Gopalakrishnan returns to more than once: humanoid robots carry a psychological burden gripper-arms never had. Because they look human, people expect them to act smart, and forgive them less when they fumble.

"A humanoid just, like, grappling around and creating a lot of failures would be judged much harshly." (Keerthana Gopalakrishnan, 66:55)

It is a strange asymmetry. A robot arm that drops a box gets a shrug. A humanoid that drops the same box, in the same way, reads as somehow disappointing.

Looking further out, the conversation turns to a genuinely open question: should robots keep reasoning the way large language models do, by looking at the present moment and reacting over and over, or do they need a forward model, a way of predicting what happens next so they can measure their own error against reality. Gopalakrishnan declines to pick a side early. "You're not emotionally attached to one method or the other method. You're emotionally attached to the problem itself," she says (73:53). The same caution shows up in how she talks about training data. Teleoperated data is precise but ages badly as hardware changes, sensor-glove data scales better but is costly to collect, and raw human video is cheap and abundant but noisy, since no two human hands move quite the same way (77:08, 77:37). Her guess is that the field keeps using all three in some shifting mixture, rather than betting everything on one.

What makes this conversation useful, more than flashy, is its refusal to round up. Robots can now run faster than Olympic sprinters, fold laundry on request, and improvise a wave mid-conversation. They still cannot reliably fry an egg in a stranger's kitchen, because an egg offers no do-over: drop it, and the room is a mess (54:00). Somewhere between those two facts sits the real timeline for a useful robot in anyone's home, and Gopalakrishnan, who builds the brains for a living, is in no hurry to guess the date.