|
There’s a lot of promise and expectations at the intersection of artificial intelligence (AI) and robotics, and a lot of work yet to be done to safely incorporate these systems into our everyday lives. Dr. Pratap Tokekar, an Associate Professor in the Department of Computer Science at the University of Maryland, is studying the science behind what AI-driven systems can and can’t do, to help determine if and when they are reliable.
In his Evening@SMART presentation on September 3, Dr. Tokekar explained that a type of AI system called a foundation model, like OpenAI's GPT series, Google's Gemini, and Anthropic's Claude, are capable of general tasks but not reliable enough to be applied to autonomous robots. Dr. Tokekar says a reliable AI model is adaptable, predictable, consistent, transparent about what it doesn't know, and capable of doing what it’s intended to do. He and his team at the University of Maryland’s (UMD) Robotics Algorithms & Autonomous Systems (RAAS) Lab are working on an alternative model to meet these requirements.
The team is working on a Vision-Language-Action model that uses vision in addition to language in order to produce actions, and uses behavior cloning, or imitation learning, to learn how to complete tasks. Once it’s shown enough processes, it doesn’t need to be shown as many new ones when learning how to complete a new task. This type of model needs quality assurance checks on its outputs to ensure both the system and the system it’s learning from are trustworthy. Several methods are used to ensure reliability.
“Tell me what you don’t know”
 Dr. Tokekar utilizes a conformal prediction framework to analyze the outputs of a system’s object detection capabilities. This involves giving the system a list of labels that is guaranteed to include the correct ones with high probability. The list of output labels is adaptive; for simpler instances, the list may be short, indicating that the model is confident in its output. For harder instances, the list may be longer, indicating uncertainty. This helps Dr. Tokekar decide if he trusts the model, or if it needs additional information and data to be more reliable.
“Be consistent”
Next, Dr. Tokekar’s team deploys a risk-constrained reinforcement learning algorithm to limit the level of a system’s failure. The algorithm optimizes the system for long-term rewards and places boundaries on how much risk or unsafe behavior the system is allowed to take. This ensures that the AI consistently avoids actions that could lead to catastrophic mistakes, even if those actions could also lead to higher rewards, helping Dr. Tokekar develop safer and more reliable performances.
“Do as intended”
 Lastly, reward models are incorporated into the system to define and guide what “good” behavior looks like. This is especially important in complex tasks when correct actions aren’t obvious. Reward models can capture human preferences, which helps the AI evaluate and prioritize its actions. Dr. Tokekar suggests training a separate network to determine what the “reward” will be. His team uses reward models to improve both the performance and usefulness of their systems.
Much of Dr. Tokekar’s work is through the RAAS Lab. The lab’s team designs algorithms and builds systems to enable teams of robots to make decisions on their own, without outside intervention.
Next Evening@SMART
The MATRIX Lab’s next Evening@SMART presentation will be on Wednesday, October 1. Dr. Miao Yu, a professor in the Department of Mechanical Engineering at UMD, will be discussing her efforts to revolutionize the shellfish aquaculture industry. Outdated technology and tools mean farmers do not have the necessary information for strategically choosing planting locations, leading to high crop mortality and reduced yields. Dr. Yu will be sharing her team’s efforts to establish smart and sustainable farming practices. REGISTER NOW
About the RAAS Lab
The Robotics Algorithms & Autonomous Systems (RAAS) Lab is part of the University of Maryland. Previously, the lab was located at Virginia Tech. The lab’s team designs algorithms and builds systems to enable teams of robots to act as sensing agents. RAAS research is at the intersection of theory and systems and is motivated by real-world applications to environmental monitoring, infrastructure inspection, and precision agriculture.
The lab is affiliated with the Department of Computer Science, UMIACS, and the Maryland Robotics Center at the University of Maryland. See more about the lab’s research here.
Related Articles:
MATRIX Interns Overcome Setbacks and Succeed
MATRIX Lab Hires Assistant Director for Research Development
MATRIX Lab Hiring Research Development Director
Clark School Seed Grant Awards Announced
MATRIX-Affiliated Faculty Solving Challenges From Sea to Space
MATRIX-Affiliated Faculty Solving Tomorrow's Challenges Today
Six Clark School Faculty Receive 2024 DURIP Awards
UMD's 40th Annual Convocation Honors Engineering Staff, Faculty
Maryland Engineers Awarded Grants to Address Humanity's Grand Challenges
Department Celebrates Achievements During Annual ‘Welcome Back’ Luncheon and Awards Ceremony
September 10, 2025
|
