Research

Proposed an approach to make use of general domain knowledge to enable sample-efficient behavior cloning. Demonstrated the effectiveness and robustness of our approach in continuous environments with discrete and continuous action spaces with very few demonstrations.

Formally defined a cognitive architecture. Showed how we can bootstrap its rules with a large language model and minimal human input. Collected a set of human preferences in the real world. Showed how the architecture we proposed can adapt to those preferences in one shot.

Proposed an agent framework that combines LLMs with customized cognitive architecture. Demonstrated how it can learn to perform various kitchen tasks from bootstrapping. Show that, when applied to new environments, it requires significantly fewer tokens than querying LLM for actions.

Showed that feature distance distributions can reflect trajectory hypotheses quality. Proposed a self quality evaluation metric SQE based on two-class Gaussian mixture model, which can primarily fulfill the self-evaluation desire. Tested the effectiveness of our method on various data sets and note its drawbacks.