I'm interested in how models learn through interaction and how to make training and inference more efficient. At Google, I'm developing learned optimizations for TPU compilers. My personal research explores learning in simulated environments and its transfer to physical robots.
When does simulation data stop helping and start erasing real physics? Rebalancing real, augmented, and simulated data when fine-tuning a VLA on my SO-101 raised pick-and-place success from 8% to 84%, without collecting more data. This post covers a six-way ablation of real / aug / sim batch composition.
Building on NVIDIA’s Voyager, this project explores how hierarchical memory and a lightweight, RL-inspired policy search implemented with DSPy can improve an embodied agent's behavior. The learned policies improved success rates by ~20% while reducing inference cost.
