Adam Haroon
Trustworthy perception and control for autonomous systems.
I am an undergraduate researcher at Iowa State University working on assured autonomy across the stack. My research asks one question twice: can we trust an autonomous system? I answer it on the sensing side by diagnosing hidden failure modes of learned 3D perception, and on the acting side by building reinforcement-learning controllers with run-time safety guarantees. This work spans 17 papers with collaborators at the U.S. Naval Research Laboratory, DEVCOM Soldier Center, and Iowa State's Virtual Reality Applications Center, and has been supported by NASA, Boeing, and the Navy Engineering Analytics Program.
One question, asked twice
Can we trust an autonomous system? I study this on both ends of the stack, and simulation is where I build and stress-test the answer. Both ends work toward one goal: assured autonomy.
Perception reliability
Can it trust what it sees?
Virtual sensors · synthetic-data benchmarks · hidden failure modes of learned perception
Control safety
Can we trust what it does?
Multi-agent RL · run-time assurance shields · provable safety & efficiency guarantees
Three pillars, one principle
Find the hidden failure, then guarantee against it.
Perception Reliability
Where learned 3D perception silently breaks: physics-faithful virtual sensors, shortcut diagnosis, and physics-based repair for fringe projection profilometry.
Control Safety
Run-time assurance and multi-agent reinforcement learning for aerial autonomy: safety as a constraint the policy cannot violate, not a reward it can trade away.
Closing the Loop
Connecting honest perception to certified control: vision-language alignment for navigation and preference-based safe offline reinforcement learning.
Highlights
Learning when to act: jointly learning control and sampling time under a Lyapunov run-time assurance shield, achieving 3.51× fewer control updates than classical methods without giving up the safety certificate.
NeurIPS 2026 · under reviewPhysics-faithful modeling of fringe projection profilometry inside NVIDIA Isaac Sim: controllable ground truth for benchmarking learned perception before it ever touches hardware.
IEEE Sensors Journal · in pressA long-range FPP network learned to lean on a shape prior instead of the signal. We diagnose the shortcut and repair it by building fixed differentiable physics back into the pipeline.
Elsevier Optics · under reviewResearch output
Venues
NeurIPS · ICLR · IEEE Sensors Journal (Q1) · IEEE Aerospace · AIAA SciTech · ACC · SPIE (Photonics West, DCS) · Elsevier Optics · CoRL / RSS workshops · SIAM SDM
Support
NASA Iowa Space Grant · Navy Engineering Analytics Program · Boeing Undergraduate Research Fellowship · SPIE Photonics West travel grant