Associate professor in computer science and software engineering earns $561K NSF award to advance AI-powered robotic perception
Published: Aug 27, 2026 1:55 PM
By Joe McAdory
Associate Professor Sathya Aakur, right, with graduate students Joe Lin and Zhou Chen, is developing artificial intelligence methods that help intelligent embodied agents, aka robots, understand what objects can do, not just what they are.
Sathya Aakur is developing artificial intelligence (AI) methods that help intelligent embodied agents, aka robots, understand what objects can do, not just what they are.
“Let’s say you want to scoop peanut butter out of a jar,” said Aakur, associate professor in the Department of Computer Science and Software Engineering and director of the Open-ended Reasoning and Knowledge Acquisition Laboratory. “Typically, you'd assume that you take a spoon and put it in there. What if there’s no spoon in that area? But there is a butter knife. Can a robot still use a butter knife to scoop it out?”
Probably, though not ideal. Do robots know this? That’s what Aakur wants to find out.
Supported by a $561,000 National Science Foundation award, Aakur’s research, “Principle-Driven Embodied Vision: Autonomous Radiance Manifold Exploration for Functional Affordance Discovery,” focuses on helping robots identify gaps in their understanding of a scene and gather the information needed to make better decisions about objects and tasks.
“What we're trying to do is within any scene or scenario, can your model or can your embodies agent position itself such that it can actively look for an object if it’s not within view?” Aakur, a 2025 Ginn Faculty Achievement Fellow, asked. “If the object is in view, can the robot look closer and get into a position where it can grasp it and also resolve any ambiguity on whether the object can be used for the desired task?”
The term “embodied agent” covers more ground than what many perceive as robots.
“It could be anything — robot dogs, humanoid robots or robotic arms,” Aakur said. “Cameras that can pan and tilt could also be embodied agents. What these models typically assume is that what you see is what you get.”
Aakur’s research targets the problem, creating a way for robots to act on their own uncertainty, moving and repositioning, instead of just detecting the object.
The project will develop persistent three-dimensional scene models that allow robots to integrate information gathered from multiple viewpoints. It also seeks to enable robots to identify when functional evidence is incomplete, determine where to look next and reason about object properties such as rigidity, containment and grasp ability.
Those capabilities will be evaluated in object retrieval and manipulation tasks that require robots to identify the object, or even the part of an object, best suited for a requested function.
Aakur's lab is testing the work on a range of physical platforms: a stationary desktop arm capable of lifting up to 5 kilograms, a quadruped robot with an arm mounted on its back that can search a room and retrieve objects, and a wheeled LocoBot outfitted with LIDAR for navigation. He also hopes to eventually test the technology on a humanoid robot, which could introduce new ways for machines to gather information about their surroundings.
“Humanoids would add another degree of freedom because you can now rotate your head, move your whole body and your neck to look at things and peek around the corner,” Aakur said. “We're probably at the ‘early 2000s Roomba state’ of humanoids at this point,” he said. “You could not see them often in homes except for a select few. Now, there’s one in almost every 20 households.”
Media Contact: , jem0040@auburn.edu, 334.844.3447