Huy Ha's Transformer Transformer designs a robot per task
The CoRL 2026 paper generates a full robot body and its controller from a target motion, and the team says the design was checked on real hardware.
Huy Ha has posted that Transformer Transformer will be presented at CoRL 2026 in Austin, with a website, video and paper all up.
The idea is robot co-design, meaning the body and the brain get optimized together instead of one being fixed. Given a target end-effector motion and a reward function, the system generates a complete robot embodiment for that task, links, joints, motors, inertial properties and the control policy. One network both designs the machine and drives it.
The hardware check
Most co-design work stops in simulation, which is why the next post matters. Shuran Song, who leads the lab behind the work, says the optimized design was also validated on real hardware, and credited the push to a reviewer.
Song also pointed at what this unlocks for data that already exists. She says dishwashing UMI data that Cheng Chi collected three years ago, using a handheld gripper rather than a robot, could be used to optimize a future robot at Sunday Robotics. Old human demos become a spec for hardware that has not been built yet.
The humanoid question
Pieter Abbeel asked the question that turns this from a neat paper into an argument about the whole field. If you run the optimizer for a single robot across enough tasks, does it converge on a humanoid? He called this example a pretty convincing stochastic gradient step in that direction.
Nobody has run that experiment yet, so nobody knows the answer. But it reframes the humanoid debate as something you could measure rather than something you pick a side on.
Danfei Xu supplied the joke of the thread, noting that his lab's upside-down mounted ALOHA arms may have been optimal all along.
Shown versus claimed. The public evidence so far is the thread, the paper page and the video, plus Song's word that a generated design was built and tested. How well the real hardware performed, and on how many tasks, is in the paper rather than the thread.
if you run for a single robot on enough tasks, does it recover humanoid?
Thanks to reviewer 2's request, we also validated the design with real-world hardware
It seems our janky upside-down mounted aloha arms are optimal all along ;)