From flood relief to farming and the frontlines, Prof. Spandan Roy is rethinking how machines learn to act in the real world.

“Even if you don’t know the system… can you still control it?” It’s not the kind of question that usually opens a talk on robotics. But for Prof. Spandan Roy, it defines everything and sets the context for his work at the Robotics Research Center at IIITH.

In theory, engineering is neat. Systems obey equations; forces can be calculated; outcomes predicted. In practice, especially in the world of flying machines, things are far less tidy. A drone in flight isn’t just governed by clean physical laws; it’s constantly negotiating wind, drag, shifting payloads, and environmental disturbances that are difficult, if not impossible, to model precisely. For someone without a traditional mechanical background, as Prof. Roy admits of himself, this gap becomes even more pronounced. Instead of trying to eliminate that uncertainty, his work embraces it. “You don’t know the dynamics but still, how can you control the system?,” he asks rhetorically.

The answer lies in designing systems that learn on the fly, adapting in real time using minimal prior knowledge, guided not by exhaustive data but by fundamental physical principles.

Drones in Motion: Precision Amid Chaos
Nowhere is this philosophy more visible than in the lab’s work with drones. What begins as a seemingly straightforward task – dropping a payload – quickly becomes a study in complexity. Imagine a drone tasked with delivering supplies onto a moving platform during a flood rescue. It must track motion, compensate for disturbances, and release the payload with pinpoint accuracy. But the real challenge comes immediately after. “When it drops, the drone suddenly shoots up… the dynamics switch completely,” explains the professor.

The system, in that instant, becomes something else entirely. Its weight changes, its balance shifts, and the rules governing its motion are rewritten mid-air. Designing a controller that can anticipate and stabilize this transition is less about precision and more about adaptability – about handling what Prof. Roy calls “switching dynamics.”

A similar tension plays out when drones carry suspended loads. A swinging payload isn’t just inefficient; it can be dangerous, especially when humans are nearby. Watching one of the lab’s demonstrations, it’s easy to see why. “If a human is just under the payload… it’s pretty nasty,” reasons Prof. Roy. The solution is not to eliminate disturbance but to respond to it- stabilizing the system even when external forces try to throw it off balance.

When Drones Do More Than Fly
As the work evolves, drones begin to take on roles beyond transportation. Attach a manipulator arm, and suddenly the drone is no longer just moving objects; it’s interacting with them. But this added capability comes at a cost. Every new component introduces instability, every movement shifts the system’s centre of gravity. In one project, the team collaborates on a novel gripper that operates without motors, inspired by the simple snap of a slap band. It’s an elegant idea, but executing it is anything but simple. “The force must be enough to grasp but not so high that it destabilizes the drone,” remarks Prof. Roy.

The challenge isn’t just mechanical; it’s deeply algorithmic. Control systems must mediate every interaction, ensuring that the act of grasping doesn’t undo the stability of flight itself.

In another experiment, the lab tackles the problem of catching objects mid-air. Here, Prof. Roy reaches for a metaphor that feels far removed from robotics: cricket. “If the hand is hard, the ball drops. If it’s soft, it absorbs the impact.” Translating that intuition into machines leads to the idea of software-driven compliance, a system where the robotic arm absorbs impact while the drone remains steady. It’s a delicate separation of roles, designed to mimic the subtle intelligence of human movement.

From Prototype to Practice
For all its technical ambition, the lab’s culture is strikingly unhierarchical. Prof. Roy describes a space where leadership shifts depending on the problem at hand. Students with hardware experience take the lead in building systems; he steps in when theory and control come into play. This model has produced not just research output, nearly thirty journal publications in six years, but a kind of continuity that compensates for the lack of long-term PhD researchers. Students train their successors, ideas carry forward, and the lab evolves as a collective.

Some of those ideas have already begun to move beyond the lab. A project on reconfigurable drones, or machines that can physically adapt to the shape and size of their payloads has spun out into a startup and found applications in collaboration with the armed forces.

In these contexts, the constraints are very different. Systems must work with minimal preparation, often under tight timelines and uncertain conditions. Here, the philosophy of adaptive control finds its most immediate test, not in controlled environments, but in situations where failure carries real consequences.

Reimagining Agriculture from the Air
Not all applications are as high-stakes, but they are no less impactful. In a collaboration on drone-based pollination, the lab turns its attention to agriculture, a domain where inefficiency is both costly and widespread. Traditional pollination methods are labour-intensive and time-bound, requiring farmers to manually agitate crops within a narrow window. The idea of using drones seems straightforward, but early attempts fall short.

What Prof. Roy’s team developed instead is a system that combines airflow with gentle physical interaction, designed to maximise pollen release without damaging the plants. The results, when they came, were unexpected. “They were hoping for 90%. We achieved 123%,” he states. Across multiple field trials, the drone-assisted method doesn’t just match manual pollination, it surpasses it.

Designing for the Unknown
Prof. Roy gestures toward the lab’s next set of challenges: robotic hands that can manipulate objects without expensive sensors, systems that infer touch through motion, machines that continue to push the boundary between knowing and doing. What ties all of this work together is not a specific application, but a way of thinking. “You don’t need to know everything about a system to make it work.”

In a field often driven by precision and prediction, it’s a quiet but radical proposition. Because the real world with all its variability, noise, and unpredictability rarely offers complete information. The question, then, is not how to eliminate uncertainty, but how to build systems that can live with it. And in that space, somewhere between control and chaos, Prof. Roy’s lab is finding its answers.