Adam HaroonAssured autonomy across the stack
Three pillars organized around one principle: find the hidden failure, then guarantee against it.
Can an autonomous system trust what it sees?
The decoding step of modern 3D perception, including fringe projection profilometry (FPP), my primary testbed, is increasingly a learned model, so it inherits every failure mode of learned models: shortcut learning, spurious correlations, distribution shift, and the sim-to-real gap. This pillar builds the instruments to expose those failures and the physics-based methods to repair them.
VIRTUS-FPP
Physics-faithful virtual FPP sensors in NVIDIA Isaac Sim: controllable ground truth for benchmarking learned perception before it touches hardware.
Elsevier Optics · under reviewShortcut diagnosis & PhiCalNet
Diagnosing shape-prior shortcuts in long-range single-shot FPP, then repairing them with a trainable backbone over fixed differentiable physics.
Elsevier Optics · under reviewDifferentiable self-calibration
Self-calibrating single-shot FPP with uncertainty propagation, plus pattern designs that make calibration identifiable by construction.
A safety guarantee is only as honest as its perception
A run-time assurance shield certifies the policy against the state it's given. If perception silently lies, the guarantee is hollow. That's why the two pillars are one research program.
If the first box is wrong, every box after it inherits the error, and the certificate certifies nothing.
Can we trust what a learned policy does?
Deep RL gives capable controllers, but with no guarantees. For systems that fly, that gap is the whole problem.
Safety
Enforce safety as a constraint the policy cannot violate, not a reward it can trade away.
Efficiency
Act only when needed: fewer control updates, less communication, same safety.
Scalability
Many agents, sharing airspace, with guarantees that hold as the team grows.
Learning when to act
Communication-efficient RL via run-time assurance: 3.51× fewer control updates under a Lyapunov shield.
IEEE Aerospace 2026 · publishedHigh-altitude balloon swarms
Distributed area coverage with multi-agent RL in stochastic stratospheric flow fields, with NRL.
IEEE Aerospace 2026 · publishedUrban air mobility
Urban Nav simulation framework and graph-encoded MARL for zero-shot scalable UAS conflict resolution.
From honest perception to certified action
The endgame is a single pipeline that carries guarantees from pixels to actions. Current work sits deliberately in the overlap between the pillars: perception that plugs into controllers which already know how to act, and policies learned safely from human and VLM preferences.
LCLA
Language-conditioned latent alignment: aligning VLM features to the latent space of a frozen privileged policy for robot navigation.
ICLR 2027 · in preparationSafe preference-based RL
State-conditioned safe offline preference-based RL: policy learning from VLM-labeled preferences with safety constraints.
Perception-to-control pipelines
Connecting perception reliability diagnostics to run-time assurance in one certified pipeline: the thesis of this research program.