Research
My main research thread is on-chip memory for accelerators. The two papers below it are course research projects from spring 2026, in power-grid control and robotics.
Course research papers
ES 145 · with Sofiyat Bello · spring 2026
Hybrid reinforcement learning and optimal power flow for smart-grid voltage control
Rooftop solar on a lightly loaded feeder pushes voltages up, and something has to decide how much generation to curtail. A PPO agent proposes curtailment for five distributed generators on a simulated 10-house feeder (three solar, one wind, one gas). Before an action is applied it runs on a pandapower AC power-flow model of the feeder; if any bus would leave 0.95–1.05 p.u., the action is replaced by an optimal-power-flow solution or, failing that, full curtailment, and the rejected case is logged for retraining.
Over three seeded 24-hour episodes (72 control steps), the hybrid controller had no voltage violations, against 91.7 % safe steps for the RL agent alone, and averaged 80.7 ms per step against 165 ms for OPF alone. In these runs every rejected action went to the full-curtailment fallback rather than to the OPF.
ES 259 · spring 2026
Learned warm starts for inverse kinematics on a UR5e
Newton–Raphson inverse kinematics is precise but depends on its starting guess, and a 6-DOF arm has several joint solutions for most poses. I trained a residual MLP (535,820 parameters) on 10 million synthetic poses from the UR5e’s product-of-exponentials forward kinematics to predict a starting configuration. It is conditioned on the robot’s current joint state as well as the target pose, which is what lets it pick the nearby solution instead of averaging between branches.
Compared with the best of five random starts, solver convergence went from 44 % to 77 % and the median number of iterations from 30 to 7, for about 0.5 ms of inference.