A conversation highlighting how human-centered autonomy shapes trust, usability, and system performance as robotics move beyond the lab into real-world use.
Madison Clark-Turner, PhD, is a deep learning and robotics research scientist in Charles River Analytics’ Sensing, Perception, and Applied Robotics division. His work focuses on how autonomous systems perceive, communicate with, and learn from people in real-world, high‑stakes environments.
Working at the boundary between research and deployment, his projects repeatedly return to the same challenge: balancing technical performance with usability, trust, and the right level of human control.
In this Q&A, Clark‑Turner reflects on how his thinking has evolved over time, what he’s learned when real-world use exposes flawed assumptions, and the principles shaping systems as they move closer to operational settings.
“Robotics gave me a practical lens for thinking about augmentation, to focus on how AI might extend what a person can perceive, decide, or do in the world.”
Q: When you look back on the moments that set your career in motion, what experiences most shaped the questions you care about, and why did robotics feel like the right place to pursue them?
A: My path into robotics wasn’t especially direct. I started out on a biology track in college, and over time I found myself spending more and more time working with computers. That eventually pulled me toward artificial intelligence and machine learning, and robotics ended up being the place where those interests came together in a tangible way.
I’ve always been drawn to science fiction and intelligent machines, but what really anchored my interest was the chance to use technology in ways that actually help people. Robotics gave me a practical lens for thinking about augmentation. Instead of optimizing systems in isolation, I could focus on how AI might extend what a single person can perceive, decide, or physically do in the world.
Q: Your work often sits at the intersection of technical capability and human needs. How do you decide what “success” looks like for a system people have to trust and rely on?
A: Success can’t just be about performance numbers. A system might look great in a lab environment and still struggle in practice if people don’t understand it, don’t trust it, or don’t feel comfortable using it when conditions are messy or unpredictable.
For me, success also shows up in how people think about the system. Do they have a clear sense of what it can do, where it falls short, and how it’s likely to behave when something unexpected happens? When users start to see the system as a collaborator rather than a black box, that’s often a stronger indicator of success than any single technical metric.
“When users start to see the system as a collaborator rather than a black box, that’s often a stronger indicator of success than any single technical metric.”
Q: There’s a lot of concern about AI replacing human workers. How do you think about automation versus augmentation in your own work?
A: A lot of it comes down to where decision-making authority sits. Humans are very good at dealing with novelty and ambiguity, especially when there isn’t a clear precedent. Most AI systems, on the other hand, are built around learning from past experience.
Autonomy tends to shine in structured environments where conditions don’t change much. The real world rarely looks like that. In unstructured settings, I think it’s important to keep humans firmly in the role of decision‑maker, with robots providing support through added sensing, physical capability, or endurance. The goal isn’t to take people out of the loop, but to help them do their work better.
“The goal isn’t to take people out of the loop, but to help them do their work better.”
Q: Can you share a project where an early assumption turned out to be wrong and what that taught you about iteration and real‑world constraints?
A: One example that stands out comes from our work on gesture recognition. Early on, I assumed we’d need a fairly large and complex model because gestures can look similar and are differentiated in how they play out over time. We built something that worked, but it also introduced challenges we hadn’t fully anticipated, like added latency and model rigidity when trying to capture how different people performed the same gesture.
What we eventually realized was that much of the information we cared about could be captured with simpler, lower‑complexity models that relied more on logic for higher-level inference . Those approaches were faster, easier to reason about, and often more robust. It was a good reminder of a lesson I keep relearning: start simple, and only add complexity when the problem really calls for it.
“Here’s a lesson I keep relearning: start simple, and only add complexity when the problem really calls for it.”
Q: You’ve studied how robots learn from people. What have you learned about the human side of that interaction that engineers often underestimate?
A: Expectations matter a lot. People come to a system with assumptions shaped by prior experience, the system’s form factor, and sometimes even pop culture. When a system doesn’t live up to those expectations, trust can drop pretty quickly.
We’ve found that being explicit about both capabilities and limitations makes a real difference. Instead of failing silently or returning a generic error, the system should explain what it understood, where it’s uncertain, and what it needs from the user. That kind of transparency helps reset expectations and makes collaboration feel more natural.
Q: Many challenges in robotics are about context rather than raw perception. How do you translate messy, real‑world environments into something a robot can act on without losing nuance?
A: Human communication goes far beyond words. Tone, gesture, posture, and environmental context all carry meaning. If someone says “pick that up” while pointing, the system has to interpret not just the language but also what’s happening around the person.
Context also builds over time. What task is already underway? What just happened? What would logically come next? Our systems need to build on that history instead of treating each instruction as an isolated command. Without that, it’s hard for robots to function outside of scripted or highly controlled environments.
“Without that [context], it’s hard for robots to function outside of scripted or highly controlled environments.”
Q: The goal of a multimodal human-robot interface is to bridge gaps between how people communicate and what robots can interpret. Can you walk through a moment that shows why that gap matters?
A: Ambiguity is a good example. If someone says “hey, can you grab that wrench” and there are several wrenches in view, the system shouldn’t guess. Instead, the system we’re developing recognizes the ambiguity and asks a clarifying question about which wrench the person means before doing anything.
It’s also built to communicate its limitations. If something is out of reach or there isn’t enough information to complete a task, it says so directly. That kind of honesty helps people understand the system’s boundaries and avoids a lot of frustration that comes from unexplained failures.
Q: Your work includes applications in time‑critical contexts like medical triage. How do you balance accuracy, usability, and transparency when the stakes are high?
A: In high‑stress situations, clarity and confidence become especially important. We pay close attention to where the system has high confidence and where uncertainty starts to creep in, whether that’s because of noise, ambiguity, or missing information.
When confidence drops below an acceptable threshold, the system needs to say that clearly and, in some cases, hand control back to the human. In life‑critical scenarios, I doubt we will ever be at the point where decision‑making should be fully delegated to machines. Human oversight remains an intentional and necessary safety boundary.
“In life‑critical scenarios, we’re not at a point where decision‑making should be fully delegated to machines. Human oversight remains an intentional and necessary safety boundary.”
Q: Looking ahead, what do you hope human-robot collaboration will make easier or safer in the next few years?
A: In the near term, I see human-robot interaction (HRI) systems making it easier for people to control very complex platforms, such as articulated robot arms or multi‑agent systems, using more natural forms of communication. That’s especially useful in unstructured manufacturing, repair, and maintenance environments where tasks change frequently.
Further down the line, I’m excited about applications in assistive care and medical robotics. Across all of these domains, the goal is the same: to build systems people understand, trust, and rely on. When technology extends human capability without overstepping human judgment, that’s where it delivers the most value.
“When technology extends human capability without overstepping human judgment, that’s where it delivers the most value.”
A few of Madison’s publications and projects
Leveraging Temporal Reasoning for Policy Selection in Learning from Demonstration
PAL/Socius: Human-robot interface to build trust and interpret speech, gestures, and gaze
MoveIt Pro Works Here – Webinar
POINTER: A robotic medical triage system that aids medics during a mass casualty incident
ASIMOV: A multicamera system that ensures safer operations using machine learning
