Mixing real and simulated data for VLA fine-tuning
A six-way data-mixture ablation on an SO-101 arm.
Pick-and-place success increased from 8% to 84%.
Nicolas Basile
ML engineer at Google, developing learned compiler optimizations for TPUs using graph neural networks and reinforcement learning. Previously, I worked on model post-training and inference optimization at VMware, including contributions to vLLM.
My personal research focuses on embodied AI and robotics. Currently exploring robotics foundation models, reproducing and fine-tuning the latest VLA and RL research in simulation (MuJoCo/MJX), and running small-scale real robot experiments.
Selected writing
Independent researchMemory and policy optimization for long-horizon agents
Extending a Voyager-style agent with hierarchical memory and DSPy-driven optimization for long-horizon tasks in Minecraft.
Read experiment : Memory and policy optimization for long-horizon agentsSelected experience
Software Engineer
Learned TPU compiler optimizations with GNNs and reinforcement learning; low-precision execution and numerical debugging.
VMware
Feb 2022 – Feb 2024Machine Learning Engineer
Model post-training for tool use, vLLM kernel contributions, and distributed inference optimization.
University of Southern California
B.S. Computer Science · B.S. Business Administration · 2021