One student’s coding frustration leads to faster computing for everyone

Gokul Puthumanaillam was developing a machine learning model to enable robots to follow any given trajectory. At the time, he was off campus, did not have access to a good computer and had to wait a long time for his model to generate results. His impatience led him to investigate the problem, develop a method to address it, and open-sourced the code so others could benefit from the same speedup.

“We’ve tested it on more than a dozen different models and found it’s scalable to any diffusion model. It works with marine robots, drones…anything that is learning to generate and follow a trajectory,” said Puthumanaillam, one of Melkior Ornik’s Ph.D. students in The Grainger College of Engineering, University of Illinois Urbana-Champaign in the Department of Aerospace Engineering.

His work earned him the Outstanding Student Paper award at the conference Robotics: Science and Systems 2026.

Puthumanaillam works with robot learning, an area of AI dealing with teaching robots how to perform actions in a smart way, a way that’s close to how humans learn to perform actions.

“Our end goal is to be able to deploy robots in settings that are hard for humans to reach but that robots can and to complete tasks autonomously, so for something like a rescue mission in a forest fire,” he said.

The area Gokul was working in uses a class of generative AI models called diffusion models to plan how robots move. Given a starting point, a destination and obstacles in the environment, these models can generate a trajectory that allows the robot to reach its goal without collisions.

But diffusion models do not produce that trajectory all at once. 

They begin with a noisy, essentially random trajectory and gradually refine it through dozens of denoising steps until a viable path emerges. At every step, the model must perform another expensive computation, making the process powerful but often too slow for robots that need to plan and react in real time.

Gokul’s work focused on eliminating much of that repeated computation without changing the trajectory the robot ultimately follows.

“I realized that denoising didn’t need to happen at every step, so I thought, why not just try skipping a few denoising steps? We figured out a way to cache the important bits of this denoising and just reuse the cache rather than recompute the cache with every step. When there is a reuse happening, it takes from the cache, and when there is a recompute happening, it writes to the cache.

“We showed that using the plugin saves a lot of computing time. By speeding up the diffusion models by over 4.4 times, it reduces wall clock time by 78%.”

He said this new way to cache diffusion models is something that has not been explored in robotics before. This is the first time a research paper has shown that it is possible.