Kumara Kahatapitiya

Google (United States), Stony Brook University

Papers

Total Citations

H-Index

About

Kumara Kahatapitiya is a researcher at the forefront of integrating sequential modeling and vision-language reasoning for robotics and visual understanding. His work centers on developing novel Transformer architectures that bridge memory, sequence modeling, and multimodal learning. Kahatapitiya’s most impactful contribution is the **Token Turing Machine (TTM)** (2023, 13 citations), which reimagines the Neural Turing Machine as a sequential, autoregressive Transformer with an external memory of tokens, enabling robust real-world visual understanding. He also introduced **StARformer** (2022, 10 citations), a Transformer that leverages state-action-reward representations to reframe reinforcement learning as a sequence modeling task, significantly advancing robot learning. Most recently, his work on **LLaRA** (2024) tackles the critical challenge of data scarcity in robotics by supercharging robot learning data for Vision-Language-Action (VLA) models, demonstrating how to adapt pretrained VLMs for control with limited demonstrations. Through these innovations, Kahatapitiya is shaping how machines learn from sequential experiences, making him a rising voice in the intersection of Transformers, memory, and embodied AI.