Two recently published papers from Oregon State University's Collaborative Robotics and Intelligent Systems Institute (CoRIS) bridge the gap between academic research and industry. One paper reported a 20% reduction in energy consumption for robot locomotion, while the other details an autonomous pruning system that could help the agriculture industry address chronic labor shortages.
Both are connected to the Dynamic Robotics and AI Lab (DRAIL) and co-authored by Stefan Lee, DRAIL co-director and holder of the Brent and Elaine Leback Endowed Engineering Professorship. Lee’s focus is on developing robots that can see, talk, and act efficiently in real-world environments.
"For a long time, robotics research lived on a table, plugged in, in a lab or factory setting," Lee said. "But more and more robots are untethered from those conditions, out in the real world. As that happens, we need to think about how we can make these robots operate 24 hours a day, 365 days a year. Because that’s the real promise of automation.”
Making robots work longer
Battery life is a constraint shared by all untethered robots. The Unitree Go2 quadruped, for example, can operate for one to four hours per charge, yet requires one to two hours to recharge. That limits the robot’s usefulness in real-world deployments.
Past attempts to improve a robot’s energy efficiency often focused on reinforcement learning with manually tuned parameters that made a trade-off between task performance and energy consumption. In this paradigm, operators must manually tune each model of robot through trial-and-error to find the best compromise.
professor and co-director of the Dynamic Robotics and AI Lab
Blue Primary, Yellow Secondary
Lee and his co-authors at DRAIL solved this with PEGrad, a machine learning technique that eliminates manual parameter tuning.
"Rather than tuning a trade-off between energy expenditure and task performance, we specify an ordering of objectives," Lee explained. This helps a robot learn behaviors that improve efficiency without reducing task success.
In simulations, PEGrad achieved a 64% reduction in motor torque while maintaining task performance. When deployed on a real Unitree Go2 quadruped, that reduction in torque translated to a roughly 20% reduction in the current draw from the robot’s battery. Researchers at CoRIS conducted tests on concrete, grass, and commercial flooring to prove the approach applies to various real-world conditions.
Beyond extending battery life, the technique can offer secondary benefits relevant to deployment costs. "A robot that moves slower and less jerkily also breaks less," Lee said. PEGrad’s elimination of manual per-robot, per-task tuning could also save time when optimizing for efficiency across heterogeneous robot fleets.
Addressing agricultural labor challenges
DRAIL researchers often work side-by-side with researchers in other CoRIS labs and have ample opportunity to collaborate. One such collaboration between Lee and Cindy Grimm, director of Oregon State’s Robotics Graduate Program, is bearing fruit in apple orchards, where autonomous pruning looks to mitigate the increasing shortage of seasonal labor.
Pruning a commercial orchard is quite different from the relaxation of pruning a home garden. It requires both precision and speed. Human pruners average one cut per second and make 10 to 50 cuts per tree; a pace robots struggle to match.
Many attempts to improve robotic pruners have relied on advanced sensors and 3D reconstructions of tree geometry to plan cuts. However, depth-sensing technologies often perform poorly in bright sunlight and may fail to capture thin branches accurately. The 3D reconstruction is also computationally intensive.
Lee, Grimm, and robotics master’s student Abhinav Jain co-authored a paper, published at the 2025 IEEE International Conference on Robotics and Automation, that describes a method for bypassing 3D reconstruction. Instead, a camera mounted on the wrist of a Universal Robotics UR5 robot arm tracks the movement of branches within its field of view. The motion patterns are used to navigate the cutting arm through cluttered tree environments.
Erasing barriers between academia and industry
Lee co-directs DRAIL with Jonathan Hurst, co-founder of Agility Robotics, Professor Alan Fern, and Associate Professor Fuxin Li. The lab collaborates with Agility and has access to the company’s latest robots, including DIGIT, the world’s first commercially deployed humanoid robot.
The lab also has access to the previously mentioned Go2 quadruped and UR5 robotic arm, as well as Agility’s Cassie bipedal robot and BlueROV autonomous underwater drones, among many others.
DRAIL’s industry engagement includes a part-time Ph.D. option that maintains ongoing connections between academic research and industry. Ph.D. student Akhil Perincherry, for example, works as a senior perception engineer at Ford while pursuing his Ph.D.
Graduates have found research and engineering positions across the tech industry. Several have moved into hardware or software engineering roles at Agility, which operates a humanoid robotic factory, RoboFab, in Salem, Oregon. Others have found work at robotics start-ups like Sanctuary AI, which pairs humanoid robots with artificial intelligence, or at companies leading the space industry, such as Honeybee Robotics (today owned by Blue Origin).
These placements reflect the applicability of skills learned in the lab. While many graduates go on to work with physical robots, the computer vision, natural language processing, and reinforcement learning techniques that students learn at DRAIL can be applied to both hardware and software challenges facing AI/ML researchers.
"With AI’s broad adoption across industries, a lot of the skills are transferable to other use cases," Lee said.
