In one Google DeepMind demonstration, a humanoid robot is in the middle of a task when someone walks up and puts a basket over its head. The robot could keep working blind. Keerthana Gopalakrishnan, research lead for Gemini Robotics at Google DeepMind, said the right response is different: the robot should notice that "something is obstructing my vision" and ask the person to take the basket off.
She used the "very funny demo" to explain how her team thinks about safety in robots. She discussed it on an episode of The Cognitive Revolution published on October 3, 2026, about Gemini Robotics 2, DeepMind's robot-model release.
Safety as something a robot has to be able to do
The question came after the conversation turned to a brief robot massage at the World Artificial Intelligence Conference (WAIC) in Shanghai. The discussion then raised the possibility that research progress might all work out, so that alignment and safety questions, rather than capability, end up as the bottleneck on when robots are deployed in homes.
Gopalakrishnan rejected the idea that the two pull against each other. "I think of safety as like a capability," she said. In her view, "people are not going to use an unsafe robot and unsafe agents," and the most useful agents will be ones that know what they need to do and are not "overriding human instructions and causing chaos or breaking laws." She said safety should be "front and center" in the design of AI systems and robots. She did not say when a household robot might arrive.
Three kinds of robot safety
Gopalakrishnan described three kinds of safety, two of which she sees as specific to robotics.
The first is operational safety: making sure a machine does not hurt people through clumsiness or instability. This has nothing to do with bad intentions. A humanoid is "not doing a vile plot to take over your house," she said, "but it's just, if it falls, it can still be quite dangerous." She said this kind of research is very specific to robotics and is not yet a big topic in the wider AI field.
The second is goal-level safety, closer to what AI safety usually means: making sure the robot does what a person asks "but without breaking guardrails."
The third is safety when sensors fail, which brings the basket back in. A robot that relies on cameras has to recognize when it can no longer see properly and respond sensibly instead of carrying on.
She stressed that safety is "a full system thing," running from the robot's mechanical design up to the high-level "brain," which she called the ER. In Gemini Robotics 2, that part is a reasoning model, Gemini Robotics ER 2 (ER stands for embodied reasoning). As AI grows more capable, she said, developers should "spend more flops thinking about safety." Flops are units of computing work, so she was arguing for devoting more computation to safety.
The published model card for Gemini Robotics ER 2 takes a similarly cautious line. It requires users to use discretion before using the models in production, commercial or public environments. It also tells them not to use the models for safety-critical applications, such as healthcare, transportation or other areas where a malfunction could reasonably foreseeably lead to death, personal injury or property damage.
Why humanoids face a higher bar
The basket stunt led to a broader point about people provoking machines. The discussion noted that people have gone out of their way to cause trouble for Waymo's driverless cars. Gopalakrishnan said she was "very surprised" by how people behave around humanoids.
She described having "two brains." Her researcher brain knows the robots are machines. Her other brain responds to how human they look, especially when they talk. "It's a bit duplicitous," she said. She is careful as a researcher, but when newcomers interact with the robots, "that boundary can blur," and she said people need to be more aware that they are dealing with machines.
She said human-likeness also raises expectations. When a robot with a simple gripper makes a mistake, nobody expects it to be smart. A humanoid "grappling around and creating a lot of failures would be judged much harshly," she said. Even when researchers explain the state of the field, "people still expect humanoids to act smarter" than robots that look nothing like people.
Gestures nobody scripted
Gopalakrishnan said Gemini Robotics 2 adds something earlier releases lacked: a human-robot interaction (HRI) component that produces natural gestures. In past demonstration videos, she said, the robot's nods were "a bit more like pre-programmed HRI." Now "the robot itself is deciding what gesture should I use as I talk to the person."
Asked where the gestures come from, she said a model prompted for human-robot interaction generates them. It is not told to make specific gestures. Instead it follows the conversation and comes up with gestures as it goes. She compared this to the action model: "No one is asking the VLA, put your hand out and then go grab the thing." VLA stands for vision-language-action model, the kind that turns what the robot sees and is told into motor commands. The discussion then turned to whether to call this behavior emergent. It was not quite emergent, the thinking went, since it was designed for in general though not gesture by gesture, so "semi-emergent" might fit.
The launch videos include interviews in which people asked the robot how it thought it had done. She said it would answer along the lines of "it was great, but maybe I made a mistake on this part of the task." By the end of filming, she said, the film crew had grown so comfortable that they would call "action" to the robot.
Her favorite moment came after the robot had been packing things in a basement and garage. Asked what the hardest part of the task had been, it named the videotape, "which was my favorite object," and said it was hard to put it in the basket instead of holding on to it. The team watching from the back burst out laughing. "Would a robot like a videotape?" she asked. She expects the lines between people and machines to blur further as humanoids get smarter and people get used to them.
Off switches and a lighter touch
The discussion also turned to emergency shutoffs, with the example of a company that put an inflated airbag in a robot's torso that disables the machine immediately if someone punctures it. Gopalakrishnan said the idea has a long history in the emergency stop, or e-stop, also called a kill switch. Most robots have a button that cuts their power.
She said e-stops come in different forms. A hard e-stop cuts all power to the body, and the robot can simply collapse, "which can then be dangerous on its own." A soft e-stop freezes the robot in place instead. She said that is often the safe choice for two-legged robots and other platforms that are not stable by themselves.
The newer progress, she said, is mostly in manipulation, where robots can increasingly control how much force they apply at the end effector, the hand or gripper that touches objects. Her example was stacking two chips: "if you press too hard, you crumble the thing." Handling delicate material takes force control and back-drivability, a joint's ability to give way when pushed instead of rigidly holding its position. She said sensing has improved enough that a robot can read the forces at its end effector, so its models "can take much better decisions than they would if they didn't have that information."