Control Theory for Neural Systems

Aerospace controls PhD turned computational neuroscientist · seeking research scientist roles

I’m currently open to research scientist positions in computational neuroscience, neural engineering, and control of complex dynamical systems. Reach out at adityad@uw.edu or on LinkedIn.

Download my CV

Hi, I’m Aditya. I’m a control theorist working on closed-loop control of neural systems. I design real-time feedback controllers and estimators that read population neural activity and steer it toward physiologically meaningful targets — across optogenetic and electrical stimulation, and across recording modalities from widefield imaging to µECoG.

I received my PhD in Control Theory from the Dept. of Aeronautics & Astronautics at the University of Washington in June 2025, advised by Prof. Mesbahi, where I worked on robust control and estimation for spacecraft guided by machine-learned perception. The questions turn out to transfer directly: controllability, observability, and robust synthesis under uncertainty are the same tools whether the plant is a spacecraft with an unreliable sensor or a cortical population you can only partially observe and only partially drive. I now bring that foundation to the design of neural interfaces and stimulation therapies, as a postdoctoral researcher at UW’s Steinmetz Lab and NERD Lab.

My research interests include:

Closed-loop control of neural population dynamics
System identification for high-dimensional, partially-observed biological systems
Controllability and reachability analysis for neural stimulation
Estimation-aware planning and robust control under set-valued uncertainty

Selected Work

Closed-loop control of neural systems

Current work — Steinmetz Lab and NERD Lab, UW.

I build real-time feedback controllers that read population neural activity and steer it toward a target. The loop uses widefield calcium imaging as an online state readout and drives a steerable stimulation laser as a function of the animal’s brain state, with quantified end-to-end latency.

The control-theoretic core is system identification: fitting low-order dynamical models to stimulation-evoked cortical responses, so a controller can be designed rather than hand-tuned. On the µECoG side, I use reachability and controllability analysis to determine which cortical nodes a given stimulation site can actually drive — the same questions that govern where to sense and steer in a trajectory problem, asked of a brain instead of a spacecraft.

Real-time closed-loop optogenetic control. A PI controller drives widefield ΔF/F activity in a target ROI to a −5% set-point; across 100 trials, closed-loop (green) tracks the reference and reduces trial-to-trial variance relative to open-loop (red).

Estimation-aware planning under set-valued uncertainty

PhD work — RAIN Lab, UW Aeronautics & Astronautics.

Paper (JGCD 2025) · Code and videos

This is where the observability machinery I now apply to neural systems came from. A neural network deployed in a physical environment often behaves like a state-dependent sensor — a keypoint network’s uncertainty, for instance, depends on the illumination it happens to be in. If uncertainty depends on state, the trajectory itself becomes a design variable for estimation quality: you can plan a path that makes the system easier to estimate while still completing the task.

I model the ML uncertainty as bounded sets, define an observability condition on the resulting output tubes, and solve the optimal control problem with sequential convex programming.

Nominal planning takes the shortest path; estimation-aware planning takes the most observable one, reducing state variance during rendezvous.

With a network of agents the setup improves further, by quantifying the directions in which information is missing and solving the problem sequentially across agents.

Multi-agent case: each agent takes a complementary path, covering the directions the others leave unobserved — the bottom row shows the illumination conditions each one is exploiting.

Experimental systems and hardware

Alongside the theory, I build the systems the theory runs on: real-time, multi-threaded image- and signal-processing pipelines in Python for online neural feedback, and before that a simulation and robotics stack for testing controllers against real sensor behaviour.

Multi-point flight test of an in-house quadrotor built for trajectory-planning experiments at the RAIN Lab.

Unreal Engine simulation of a satellite in Earth orbit with a chaser camera view — Unreal Engine environment I built for vision-based rendezvous, coupling a custom dynamics engine to a learned perception pipeline.

I also built and supervised educational hardware testbeds for aerial and ground robots at the RAIN Lab, including a ROS2 ground-robot platform for testing trajectory optimization and an in-house indoor positioning system.

Selected Publications

A. Deole, N. Steinmetz, et al. “Towards Data-driven Feedback for Cortical Activity.” NeuroAI, 2025.
Z. Lu, A. Deole, et al. “Benchmarking Probabilistic Time Series Forecasting Models on Neural Activity.” NeurIPS 2025 Workshop.
A. Deole, M. Mesbahi. “Estimation-Aware Trajectory Optimization with Set-Valued Measurement Uncertainties.” Journal of Guidance, Control, and Dynamics, 2025.

See all publications →