Projects
-
Deep reinforcement learning project
A foveated ImageNet-C classifier. Goal-conditioned PPO agents commit high-resolution patches under distribution shift. The ablation tests whether prediction error as an observation feature accelerates adaptation, the core claim of the adaptive-sensing work.
-
Operational-space control in simulation via SCL, the Stanford Robotics Lab's framework. The direction is contact-rich manipulation with an adaptive-sensing readout, extending the prediction-error gate from my Arbabian Lab work into closed-loop torque control.
-
CS 153, Frontier Systems, taught by Anjney Midha and Michael Abbott. The "one-person frontier lab" project is a single artifact at the new ceiling of what one researcher can ship end-to-end with current frontier tools.