Gemini 4 is a good model for full-stack robotics
I've been impressed watching it autonomously handle hours of CAD, sysid, simulation, policy training, etc
To celebrate the boost, we let it develop the hardware and software for a highly seasonal robot
Happy Halloween!
While most jack-o-lanterns just sit on porches, the @GoogleDeepMind team decided to let AI agents design, simulate, and bring one to life. 🎃🤖
We built a physical, dancing 12 degrees of freedom robot directly from a single plain-English prompt, with no manual CAD required!
Can a robot assemble tight-clearance parts using only its proprioception? No camera, no tactile sensor, no F/T sensor, no joint torque sensing, no priors over the object geometries or poses: just joint position readings.
It actually can! Blind policies trained in simulation
Actuator networks offer a data-driven alternative to traditional sysid, but until now, they have been difficult to train for humanoids.
In Contact-UAN, @nolan_fey@khai_ngx uncover that the practical bottleneck is state estimation - and that this is solvable with careful design!
(1/9) Same rewards. Less domain randomization. Higher jumps. Stable landings.
Just by improving the actuator model.
Introducing Contact-UAN: a residual world model for humanoids, learned from proprioception. No torque sensing or motion capture.
Project website:
Turn your favorite whole-body tracker into a compliant one with the recipe from @venki1899@michl_wang ! Should be faster, easier, and more performant than training from scratch
As humanoid robots become part of our everyday lives, they need to do more than just perform impressive motions. They should be safe, fun to interact with, and respond naturally to unexpected contact, without sacrificing performance.
Our latest work, Compliant GMT (C-GMT),
