When a humanoid robot steps forward, the forces on its knee depend on what its hip and ankle are doing at that same moment. Physics is full of these multiplied relationships. Yet the neural networks that control many robots are built on a design borrowed from image recognition, which does not multiply anything directly. Researchers at the University of Michigan say fixing that single mismatch can make robots noticeably steadier.
The team’s new component, PRISM, was described in a Michigan Engineering announcement on November 10, 2026 and presented at the 10th Conference on Robot Learning (CoRL 2026) in Austin, Texas.
The problem with borrowed networks
An image classifier maps a picture to a label and leaves the picture unchanged. A robot controller is different: its actions change its own physical state, which changes what its sensors report next. According to the university, motion and goals involve multiplicative relationships among variables. Forces at one joint depend on the positions and velocities of other joints, and mechanical power is force times motion.
Standard multilayer perceptrons, or MLPs, do not explicitly multiply variables. They approximate those interactions through simpler operations, which uses up network capacity. The researchers argue that this is why simply making networks bigger is not the best path to better motor control.
What PRISM does differently
PRISM stands for Polynomial Representations for Interaction-Structured Motor Control. It is an input module that learns multiplicative interactions among sensor readings, past actions and task commands, such as joint positions, velocities and a desired walking speed. Engineers do not have to write down physical equations or pick which variables to combine; the module learns that from data.
It replaces only the first stage of processing, so it can slot into existing robot learning systems without changing training objectives, sensing requirements or controllers. Lead author Seung Hyun Lee, a doctoral student, said many useful clues about physical interaction already sit in a robot’s observations, and PRISM gives the controller a compact way to use them.
The numbers
In simulated humanoid walking, PRISM completed 92.5% of 24-second trials without a fall, compared with 51% for a conventional MLP. A conventional network enlarged by about 43% to match PRISM’s size managed only 52.3%.
In a simulated cube-stacking test where the arm was deliberately made to drop the cube, PRISM recovered and finished the task in 52.3% of trials, versus 40.9% for a conventional imitation-learning policy. On real hardware, an SO-101 robotic arm picking up a roll of tape succeeded in 94% of 50 attempts with PRISM, against 68% for the conventional approach, using identical training demonstrations. The team also reports gains for a large vision-language-action model.
Handling conditions it never trained on
Changes in friction, mass and payload without retraining are a common way robots fail. Across six such conditions, PRISM-controlled humanoids completed 92.3% of flat-terrain trials versus 78.3% for a parameter-matched MLP. On uneven terrain the figures were 87.1% versus 80%.
The study also reports that PRISM reduced contact forces quickly after touching an object in simulation, without force feedback sensors or a hand-designed compliance controller. Internally, its representations made quantities such as joint power, slipping and contact forces more predictable than those learned by conventional networks.
What it means
For robot developers, the appeal is efficiency. Better behavior without extra sensors or a larger model can lower cost and computing needs. Because the module drops into existing pipelines, teams could test it on current projects. For students and researchers, it is a reminder that how a network represents a problem can matter more than how big it is.
Caveats apply. Most results come from simulation, with one real-arm test, and the work is a research study rather than a product. Senior author Stella Yu, a professor of computer science and engineering, said the results make a strong case for using PRISM in place of conventional MLP input layers in robot policy learning, which the wider community will now test.
Conclusion
PRISM suggests a modest but powerful idea: give a robot’s network the physics-shaped structure of its own body, and it needs less brute force. The study was supported by the National Science Foundation, and its next test will be whether other labs reproduce the gains on more robots.