
Sample-Efficient Reinforcement Learning for Policy Improvement via Flow-Reversed Action Anchors
We developed AnchorQ to steer a pretrained robot policy by learning which variations on a familiar motion work best.
AI · Robotics · Embodied Intelligence
Applied scientist and technical lead building learning systems for dexterous robot manipulation.
I work across imitation learning, reinforcement learning, perception, and deployed robotics—turning research into systems that operate at industrial scale.


We developed AnchorQ to steer a pretrained robot policy by learning which variations on a familiar motion work best.
A robot that autonomously solves a Rubik's cube.

We used a reasoning agent to review failed robot trials and suggest specific movements for the policy to try during training.

We trained a robot to describe what it has seen and use that description to guide its next movement.

We trained a robot to predict how its hand, a tool, and the object it acts on will move together.