Robots Just Got a Brain That Works Across Bodies

For years, robots that walk and grab things stayed stuck in the demo phase. Cool video, no real work. That gap is starting to close. Google DeepMind put out a robot brain that runs across different bodies. A new humanoid can feel where you touch it. And a startup with no roboticists just raised $70 million to figure out how people should talk to machines at all. Let me catch you up on what actually moved.

One brain, many bodies

The old problem was simple and brutal: teach a robot a skill, and it stayed locked to that one robot. Move to a new body and you started over. Gemini Robotics 2 chips away at that. DeepMind says the same model checkpoint controlled Apptronik's Apollo humanoid, its own dual-arm setup, and a Franka Duo platform. One brain, different hardware.

It also moved past tabletop tasks. The Verge reports the model now handles "whole-body motions" from feet to fingertips, so a humanoid can walk to a shelf, bend down, and place a watering can where you asked. The on-device version adapts to a brand-new robot body in a few hours, usually with fewer than 200 examples.

For a leader, that changes the buy-versus-build math. You may not need to bet on one robot vendor and pray. The intelligence layer is starting to travel.

The hands finally got fingers

Grabbing a block is easy. Sealing a ziplock bag is not. Gemini Robotics 2 can now drive a five-fingered, 22 degree-of-freedom hand to tie knots and seal bags, and also run a plain two-finger gripper for tight packing. Same model, different end effectors.

The reasoning side got sharper too. Gemini Robotics ER 2 watches live video and tracks how far along a task is, hitting 91.3% on spotting the exact frame a key event happens, like when to stop pouring. That solves a boring but real problem: knowing when a job is actually done before moving on.

DeepMind is honest about the limits. Movement speed still lags, and multi-finger dexterity is uneven across benchmarks. This is capability, not a finished product.

Skin, so the robot knows you're there

The other half of usefulness is safety near people. A humanoid called Gene.01, from Generative Bionics, showed up at AMD's conference wrapped in full-body "smart skin" that senses touch, proximity, force, and temperature. Squeeze its arm and a screen lights up at the exact spot, darker as you press harder.

It uses time-of-flight sensors to notice a hand before contact, so it can avoid a collision instead of reacting after the bump. The point is factory-floor safety, letting robots work close to humans.

DeepMind is pushing the same goal from software. ER 2 halts a humanoid when a person steps near and resumes only when the area clears. If your product ever puts a machine next to a worker, this is the part your safety and legal teams will ask about first.

Nobody knows how to talk to these things yet

Here is the gap the demos hide. Enigma raised $70 million to study a plainer question: how should a person tell a robot what to do? Co-founder Jonathan Jacobi put it well. If loading the dishwasher takes 15 minutes of explaining, people quit and do it themselves. "Right now, everyone is at that point, even with the most capable models."

Enigma opened more than 100 real robots online for anyone to control through a browser, testing text, audio, and demonstrations. The founders are not roboticists. They came out of Israel's Unit 8200. As Packy McCormick framed it, no PhD means no stuck in one path.

The lesson for you: capability and usability are two different roadmaps. A robot that can seal a bag is worthless if telling it to takes longer than doing it yourself.

The deep cut

The headline is whole-body control. The thing that changes your planning is portability. When one model runs across a humanoid, a rover, and a two-arm gripper, you stop being locked to a single hardware vendor and their release schedule. Your leverage moves up a layer.

So when a robotics pitch lands on your desk, stop asking only what this robot can do. Ask what carries over. Does the intelligence transfer to other bodies, or are you buying a dead end that needs a full retrain the day you switch suppliers? The answer tells you whether you are buying a capability or a cage.

Three questions for your team

• If we piloted a physical-automation task next year, which one is real enough to test now and which is still a demo? Name the task, not the robot.
• For any human-adjacent deployment, what does our safety bar look like, and does the vendor's proximity detection meet it? Get specifics, not a promise.
• How much time does it take a worker to instruct the machine versus do the task by hand? If instructing is slower, the project is not ready, no matter how good the demo looks.