92.5%: Michigan's PRISM improves simulated humanoid walking
A neural network component learns how sensor readings, movements and commands interact, raising success in walking and grasping tasks.
Researchers at the University of Michigan added a new part to a robot's brain that helps it understand how what it senses relates to what it does. It walked and grasped things more successfully in tests. It is still research, not a product.
Hype check
Louder than it deserves. The real change is smaller.
Our editorial opinion, based on the sources below. How we score
Your call: hype or real?
Real = it will matter in daily use. Hype = louder than it is. One vote, and it can’t be changed.
Our view and readers’ views are opinions, not statements of fact about a product.
Why it matters
Robot learning improves without changing its sensors or goals, but the idea is still research, not a product.
Watch: steadier robots, but probably years away.
The details
- PRISM was developed at the University of Michigan; lead author Seung Hyun Lee, senior author Professor Stella Yu.
- In a simulated humanoid, 92.5% of 24-second trials finished without a fall, against 51% for a conventional network.
- On a real SO-101 robot arm, picking up tape succeeded in 94% of 50 attempts, against 68% for the conventional approach.
- In cube stacking after a deliberate drop, success was 52.3% against 40.9%.
- The paper was presented at CoRL 2026; most tests were in simulation and no commercial plans are announced.
Receipts
One trusted outlet reports it. Not confirmed yet. Open the sources and check us.
Written with the help of AI tools and checked against the sources above. How we work
See a mistake? Tell us and we’ll fix it.
