Achyuth K. Vivek
Junior, Computer Science & Applied Mathematics · UC Berkeley
I'm an undergraduate researcher in the Robotic AI & Learning (RAIL) Lab at UC Berkeley, advised by Professor Sergey Levine.
I work on data-constrained reinforcement learning for robotics, specifically focusing on sample efficiency and noisy data.
Research
Reinforcement learning works well when interaction is cheap. On real robots it isn't — every trial costs wall-clock time and hardware wear, and the data you do collect is noisy and off-distribution. Most of my work comes back to that constraint: how to learn good policies from limited, imperfect real-world experience.
Right now I'm working on:
- Sample-efficient RL for robotics. Squeezing more out of each interaction, so that learning on real hardware stays practical.
- Agentic task decomposition. Using LLM agents to break long-horizon tasks into subtasks that can be learned and run in parallel — automating parts of the RL pipeline that are usually hand-designed.
- World models from noisy data. Learning predictive models that stay reliable when the training data is imperfect, off-policy, or only loosely labeled.