South Dakota Mines Professor Designs First Emotion-Driven Navigation System for AI Agents

Artificial Intelligence is everywhere—helping doctors detect diseases, guiding self-driving cars through traffic, powering the apps on our phones, and even supporting technologies that help older adults live more independently.

Despite these breakthroughs, AI still struggles with something humans do naturally: adapting when things go wrong. It can compute, calculate, and optimize, but when faced with setbacks or negative experiences, AI systems typically falter.

That’s where the work of Nirmalya Thakur, Ph.D., assistant professor in the Department of Electrical Engineering and Computer Science at South Dakota Mines, comes in.  Recognized by the IEEE Computer Society as one of the Top 30 Early Career Professionals of 2024, Thakur has developed the first emotion-driven navigation system for AI agents. His project, Emotional Detours, enables AI agents to recover from setbacks and continue efficiently toward their goals, mirroring how people overcome challenges in daily life.

The system equips an AI agent with an emotion-driven approach to navigation: a setback, such as entering an undesirable region, triggers the agent to detour to the nearest recovery point and then resume its path toward the goal. And, it’s memory updates to avoid the same setback in the future.

AI has come a long way, but one gap has remained — the ability to recover after setbacks. Emotional Detours shows, for the first time, how an AI agent can face difficulty, recover, and continue toward its goal in a way that feels more human-like.

“In real-world environments, undesirable regions may include areas such as slippery floors or poorly lit corridors, places that can pose risks to an agent’s navigation, such as slipping or slowing its overall progress,” Thakur said. “Rather than only detecting emotions, the system uses principles inspired by human emotional behavior — treating setbacks as negative experiences, seeking recovery, and then continuing toward the goal.”

Unlike many AI systems that rely on large volumes of training data, Thakur’s system operates entirely without training data, instead relying on intelligent decision logic and real-time environmental analysis, supported by an adaptation process that guides responses during navigation.

To evaluate the system, it was tested in 72 simulated environments, each with different layouts and new start and goal points for the agent.

“The system is designed to find the most efficient path to the goal,” Thakur said. “But when it experiences a setback — like entering an undesirable area — it doesn’t just force its way through. It takes a recovery step and then continues forward. This ability to adapt in real time, inspired by how people recover from negative experiences, is what makes this system stand out.”

The agent also remembers those areas, so it doesn’t return to them later.

“It reflects how people deal with challenges — when something goes wrong, we remember it and try not to repeat it,” he said. “Emotional Detours applies that same emotional principle to AI, allowing an agent to avoid the same setbacks and make steady progress toward its goal.”

Across the simulated environments, the system consistently outperformed a baseline approach, producing shorter routes, fewer entries into undesirable regions, and faster recovery.

Emotional Detours builds on Thakur’s prior research on adaptive technologies to support healthy aging, independent living, and quality of life in smart homes. This concept could one day be applied in areas such as healthcare, self-driving vehicles, and smart homes, where systems must adapt to people’s needs in real time.

While still a proof of concept, Emotional Detours demonstrates how an emotion-driven navigation system with recovery and memory can make AI agents more resilient and trustworthy. Thakur’s work could reshape how humans and machines interact, opening the door to more intuitive and empathetic technology in everyday life.

Thakur’s paper on Emotional Detours has been accepted for presentation at the IEEE IEMCON 2025 conference, scheduled to take place this month at the University of California, Berkeley.