Aastha Jhunjhunwala is a solutions architect at NVIDIA working on foundation-model development, scientific AI, and robotics. Her contributions include the Nemotron-4 language models and GR00T N1, a humanoid foundation model that translates visual observations and language instructions into physical actions.
Jhunjhunwala earned a master’s degree in chemical engineering at Carnegie Mellon University, concentrating on molecular dynamics and machine learning for drug discovery. At NVIDIA, she initially wrote about GPU-accelerated mathematical libraries before expanding into language-model pretraining, inference, and generative AI infrastructure.
She coauthored the 2024 technical reports for Nemotron-4 15B and Nemotron-4 340B, contributing to the latter’s foundation-model team. She also participated in a biochemical foundation-model collaboration involving nach0, which combines natural and chemical language.
- Cross-embodiment action decoding. In her GR00T N1 architecture explanation, Jhunjhunwala describes how a slower planning system coordinates with faster action generation, processing images, language, and robot-state information through a diffusion-transformer architecture. An embodiment-specific decoder translates shared model representations into movements appropriate to a particular robot body. She contrasts imitation learning’s dependence on expensive demonstrations with reinforcement learning’s difficulty transferring behavior from simulation into physical environments.
More recently, Jhunjhunwala contributed to Nemotron 3 Nano Omni within its evaluation, product, and legal contributor group.
