AI

Robot regret: New research helps robots make safer decisions around humans

Jeff Zehnder — Fri, 29 Aug 2025 16:28:28 +0000

Robot regret: New research helps robots make safer decisions around humans Jeff Zehnder Fri, 08/29/2025 - 10:28

Imagine for a moment that you’re in an auto factory. A robot and a human are working next to each other on the production line. The robot is busy rapidly assembling car doors while the human runs quality control, inspecting the doors for damage and making sure they come together as they should.

Robots and humans can make formidable teams in manufacturing, health care and numerous other industries. While the robot might be quicker and more effective at monotonous, repetitive tasks like assembling large auto parts, the person can excel at certain tasks that are more complex or require more dexterity.

But there can be a dark side to these robot-human interactions. People are prone to making mistakes and acting unpredictably, which can create unexpected situations that robots aren’t prepared to handle. The results can be tragic.

New and emerging research could change the way robots handle the uncertainty that comes hand-in-hand with human interactions. Morteza Lahijanian, an associate professor in ��’s Ann and H.J. Smead Department of Aerospace Engineering Sciences, develops processes that let robots make safer decisions around humans while still trying to complete their tasks efficiently.

From left, engineering professor Morteza Lahijanian and graduate student Karan Muvvala watch as a robotic arm completes a task using wooden blocks. (Credit: Casey Cass)

In a new study presented at the International Joint Conference on Artificial Intelligence in August 2025, Lahijanian and graduate students Karan Muvvala and Qi Heng Ho devised new algorithms that help robots create the best possible outcomes from their actions in situations that carry some uncertainty and risk.

“How do we go from very structured environments where there is no human, where the robots are doing everything by themselves, to unstructured environments where there are a lot of uncertainties and other agents?” Lahijanian asked.

“If you’re a robot, you have to be able to interact with others. You have to put yourself out there and take a risk and see what happens. But how do you make that decision, and how much risk do you want to tolerate?”

Similar to humans, robots have mental models that they use to make decisions. When working with a human, a robot will try to predict the person’s actions and respond accordingly. The robot is optimized for completing a task—assembling an auto part, for example—but ideally, it will also take other factors into consideration.

In the new study, the research team drew upon game theory, a mathematical concept that originated in economics, to develop the new algorithms for robots. Game theory analyzes how companies, governments and individuals make decisions in a system where other “players” are also making choices that affect the ultimate outcome.

In robotics, game theory conceptualizes a robot as being one of numerous players in a game that it’s trying to win. For a robot, “winning” is completing a task successfully—but winning is never guaranteed when there’s a human in the mix, and keeping the human safe is also a top priority.

So instead of trying to guarantee a robot will always win, the researchers proposed the concept of a robot finding an “admissible strategy.” Using such a strategy, a robot will accomplish as much of its task as possible while also minimizing any harm, including to a human.

“In choosing a strategy, you don't want the robot to seem very adversarial,” said Lahijanian. “In order to give that softness to the robot, we look at the notion of regret. Is the robot going to regret its action in the future? And in optimizing for the best action at the moment, you try to take an action that you won't regret.”

Let’s go back to the auto factory where the robot and human are working side-by-side. If the person makes mistakes or is not cooperative, using the researchers’ algorithms, a robot could take matters into its own hands. If the person is making mistakes, the robot will try to fix these without endangering the person. But if that doesn’t work, the robot could, for example, pick up what it’s working on and take it to a safer area to finish its task.

Karan Muvvala watches the robotic arm pick up a blue block. (Credit: Casey Cass)

Much like a chess champion who thinks several turns ahead about an opponent’s possible moves, a robot will try to anticipate what a person will do and stay several steps ahead of them, Lahijanian said.

But the goal is not to attempt the impossible and perfectly predict a person’s actions. Instead, the goal is to create robots that put people’s safety first.

“If you want to have collaboration between a human and a robot, the robot has to adjust itself to the human. We don't want humans to adjust themselves to the robot,” he said. “You can have a human who is a novice and doesn't know what they're doing, or you can have a human who is an expert. But as a robot, you don't know which kind of human you're going to get. So you need to have a strategy for all possible cases.”

And when robots can work safely alongside humans, they can enhance people's lives and provide real and tangible benefits to society.

As more industries embrace robots and artificial intelligence, there are many lingering questions about what AI will ultimately be capable of doing, whether it will be able to take over the jobs that people have historically done, and what that could mean for humanity. But there are upsides to robots being able to take on certain types of jobs. They could work in fields with labor shortages, such as health care for older populations, and physically challenging jobs that may take a toll on workers’ health.

Lahijanian also believes that, when they're used correctly, robots and AI can enhance human talents and expand what we're capable of doing.

"Human-robot collaboration is about combining complementary strengths: humans contribute intelligence, judgment, and flexibility, while robots offer precision, strength, and reliability," he said.

"Together, they can achieve more than either could alone, safely and efficiently."

Off

Traditional

White

�� advancing artificial intelligence research for real-world applications

Caroline Harrah — Mon, 25 Aug 2025 15:58:15 +0000

�� advancing artificial intelligence research for real-world applications Caroline Harrah Mon, 08/25/2025 - 09:58

Caroline Harrah

A modified Boston Dynamics Spot robot operates in a mine as part of an experiment leading up to the DARPA Subterranean Challenge.

�� researchers maneuver a robot through fallen rock and debris in an experimental mine tunnel in Colorado. Somewhere in the darkness ahead, potential survivors could be running out of air. The robot stops, assesses the unstable ceiling above, then continues toward areas too dangerous for human rescuers.

Other teams in the College of Engineering and Applied Sciences are developing systems that analyze satellite imagery to identify environmental changes invisible to the human eye, while computers mine years of medical records in seconds to help doctors save lives.

Meanwhile, researchers are developing computer systems that describe unfamiliar spaces in detail, guiding someone through a building by noting, for example, "stairs ahead, handrail on your right, doorway opens to a lobby beyond."

These scenarios illustrate a few of the artificial intelligence (AI) discoveries happening at ��, where researchers collaborate across disciplines to develop trustworthy systems that work alongside humans to address challenges in emergency response, space and planetary exploration, medical diagnosis, education, environmental prediction, accessibility applications and more.

When robots venture where humans cannot

Sean Humbert, professor of mechanical engineering and director of ��'s Robotics Program, oversees a lab where autonomous systems and bio-inspired robots learn to operate in places and situations where direct human engagement creates unacceptable risk. The algorithms developed by Humbert’s team must process sensor data within milliseconds, adapting to terrain that shifts without warning in contaminated areas, deep ocean trenches and collapsed structures where every decision could mean the difference between rescue and catastrophe.

Humbert, along with Christoffer Heckman and Eric Frew, faculty in computer science and aerospace engineering sciences, respectively, recently led a team of CU researchers in the DARPA Subterranean Challenge to test these systems in unmapped caves, where GPS and communication links were unavailable. Equipment failures and navigation problems in real conditions taught the team lessons impossible to learn in the lab. This iterative process of real-world testing proves essential for developing truly reliable autonomous rescue systems.

��'s first humanoid robot is being prepared to perform electric vehicle (EV) lithium-ion battery pack disassembly.

Heckman has also partnered with the Army Research Laboratory to investigate the use of vision-language models, such as those employed in ChatGPT, for building robotic contextual awareness. Building on these AI agents, Heckman’s group is developing models that help interpret statements like “go check out the backpack in the rear of the building,” which provide spatial context to robotic assistants.

Morteza Lahijanian, associate professor of aerospace engineering sciences, approaches the same domain with a fundamentally different methodology. Where some systems learn from experience, Lahijanian applies formal verification techniques to help ensure safe behavior before deployment. His team is pioneering a new AI paradigm that complements learning with logic-based reasoning, enabling autonomous systems to make decisions grounded in both intuition and logic.

In space applications, where missions are costly and communication delays are unacceptable, Lahijanian develops autonomous systems for spacecraft that must dock with satellites and space stations without human intervention. His mathematical verification methods ensure these systems can handle situations no programmer could anticipate.

Teaching robots to work with people

Human-robot collaboration requires systems that can learn from and communicate with their human teammates. Brad Hayes, associate professor of computer science, combines advances in machine learning techniques, including Explainable AI, Reinforcement Learning, and Imitation Learning, with foundational principles from robotics and psychology to develop intelligent, reliable, and adaptable autonomous collaborators. By working at the intersection of pervasive and personalized artificial intelligence, large language models, collaborative agents, and decision support, his work produces the technologies that empower human-robot teams to be greater than the sum of their parts.

Alessandro Roncone, assistant professor of computer science and associate director of the Robotics Program, designs robots that act as teammates by understanding human goals and adapting accordingly. His work addresses tricky questions: When should the robot override the human? How can it warn people about limitations without being annoying? Roncone’s research suggests that some chemists hate when robots second-guess their decisions. Others become too dependent on robotic assistance. Achieving balance requires understanding individual personalities.

The challenge of making machines understand

James H. Martin, professor of computer science and faculty fellow at the Institute of Cognitive Science, works on natural language processing systems that can capture meaning in language and exploit that knowledge in practical applications. Language is often vague and ambiguous, requiring context and background knowledge to achieve adequate interpretations. Despite advances with large language models, current approaches struggle with ambiguous language. Supported by awards from the National Institutes of Health, Martin and his colleagues have been developing systems to mine electronic medical records for temporal and causal patterns to improve clinical outcomes. As part of his work with the National Science Foundation AI Institute for Student-AI Teaming (iSAT), located here at ��, Martin and his students are contributing to systems to analyze and improve instructional discourse in K-12 STEM classrooms.

Professor Alessandro Roncone and a PhD student are working on a collaborative robot that, in the future, will assist people across complex environments, from factories to hospitals.

Maria Pacheco, assistant professor of computer science, faces her own linguistic puzzle. While LLMs are adept at interpreting and producing language, how to use them effectively for knowledge-intensive purposes remains an open question. Whenever LLMs need access to knowledge they were not trained on, or to provide transparent, faithful explanations for their decisions, they fall short. Pacheco studies how to integrate the language capabilities of LLMs with external, structured knowledge resources to develop robust language technology employed in real-world applications.

Beyond recognition: AI that truly sees

Danna Gurari, assistant professor of computer science, applies neural network models to generate spatial descriptions rather than standard object classification. Developed with input from individuals who have vision impairments, her systems provide detailed image descriptions that enable AI-generated spatial navigation. People require different types of information in various contexts. Her work addresses the challenge of providing relevant details without overwhelming users with unnecessary information.

Standardizing how we teach and visualize AI architectures

Associate professor Tom Yeh’s AI research, represented by the AI by Hand project, bridges human-computer interaction and computer vision to make deep learning more intuitive and accessible. With over 200,000 followers across social media platforms, the project has achieved global impact by demystifying complex AI concepts through hand-drawn visual explanations. Yeh invented a novel unified representation framework for neural network architectures, designed to help learners and practitioners better understand, compare and build models. By combining rigorous technical insight with human-centered design, Yeh’s research empowers diverse audiences to engage more meaningfully with the foundations of modern AI.

Embedding Earth's data to improve environmental monitoring

Satellite data offer an unprecedented window into Earth and its changing conditions, but turning this firehose of data into usable information requires new technologies. Esther Rolf, assistant professor of computer science, develops machine learning systems that coalesce satellite data into succinct, general-purpose representations of the Earth. These embeddings unlock new large-scale applications in environmental monitoring; her recent research identified previously unmonitored artisanal mining activity across Sub-Saharan Africa. Rolf pushes the frontier of geospatial AI, aligning fundamental research advances with goals of computational efficiency and accessibility.

Looking forward

As AI systems integrate into scientific, industrial and public domains, �� researchers are addressing both technical challenges, such as ensuring reliability and explainability, and social dimensions, including fairness and privacy protection. Rather than replacing human judgment, research aims to create AI that enhances human capabilities and decision-making.

Whether developing robots for dangerous rescue missions, creating educational tools that help people understand AI, or building systems that track environmental changes across the globe, �� researchers are advancing technology that amplifies human capability and safeguards human agency.

From robots navigating disaster zones to AI that supports doctors, astronauts, and students, �� researchers are creating real-world systems that redefine how artificial intelligence serves humanity. Their work spans life-saving rescue missions, space exploration, environmental monitoring, and human-AI collaboration and more. Discover how �� is leading the next wave of AI innovation.

Off

Traditional

White

Civil, environmental and architectural engineering welcomes two new faculty

Susan Glairon — Sat, 16 Aug 2025 10:57:58 +0000

Civil, environmental and architectural engineering welcomes two new faculty Susan Glairon Sat, 08/16/2025 - 04:57

Susan Glairon

Meet Assistant Professors Laura Sunberg and Zhi Li — and see why we’re so excited to have these talented scholars on our team.

Off

Traditional

White

Research institute building the AI-literate workforce of the future receives major new grant

Emily Adams — Thu, 31 Jul 2025 19:28:04 +0000

Research institute building the AI-literate workforce of the future receives major new grant Emily Adams Thu, 07/31/2025 - 13:28

Pioneering research institute led by the �� launched in 2020 to explore how classrooms could become more effective and engaging learning environments.

Off

Traditional

White

Researchers test the trustworthiness of AI—by playing sudoku

Emily Adams — Mon, 28 Jul 2025 17:15:26 +0000

Researchers test the trustworthiness of AI—by playing sudoku Emily Adams Mon, 07/28/2025 - 11:15

In a new study, a team of computer scientists and engineers from the �� created nearly 2,300 original sudoku puzzles, which require players to enter numbers into a grid following certain rules, then asked several AI tools to fill them in.

Off

Traditional

White

�� launching master’s degrees in artificial intelligence

Emily Adams — Wed, 23 Apr 2025 13:00:00 +0000

�� launching master’s degrees in artificial intelligence Emily Adams Wed, 04/23/2025 - 07:00

To meet increased workforce demand, the Department of Computer Science is introducing a Master of Science degree in AI that will be available both in-person and online through Coursera.

Off

Traditional

White

As AI explosion threatens progress on climate change, these researchers are seeking solutions

Susan Glairon — Fri, 18 Apr 2025 22:31:53 +0000

As AI explosion threatens progress on climate change, these researchers are seeking solutions Susan Glairon Fri, 04/18/2025 - 16:31

Kyri Baker, associate professor in the Department of Civil, Environmental and Architectural Engineering, and her colleague, Bri-Mathias Hodge, professor in the Department of Electrical, Computer & Energy Engineering, suggest that if future data centers are placed in the right location and equipped with energy storage technologies, they can run on 100 percent clean energy.

Off

Traditional

White

AI

Robot regret: New research helps robots make safer decisions around humans

Off

���������� advancing artificial intelligence research for real-world applications

When robots venture where humans cannot

Teaching robots to work with people

The challenge of making machines understand

Beyond recognition: AI that truly sees

Standardizing how we teach and visualize AI architectures

Embedding Earth's data to improve environmental monitoring

Looking forward

Off

Civil, environmental and architectural engineering welcomes two new faculty

Off

Research institute building the AI-literate workforce of the future receives major new grant

Off

Researchers test the trustworthiness of AI—by playing sudoku

Off

���������� launching master’s degrees in artificial intelligence

Off

As AI explosion threatens progress on climate change, these researchers are seeking solutions

Off

�� advancing artificial intelligence research for real-world applications

�� launching master’s degrees in artificial intelligence