Action Embeddings

Action embeddings represent actions as vectors in a latent space, aiming to capture their semantic meaning and impact for improved machine learning performance. Current research focuses on leveraging these embeddings within various model architectures, including diffusion models for procedure planning, and reinforcement learning algorithms for policy optimization and off-policy evaluation, particularly in scenarios with large action spaces. This work is significant because effective action embeddings enhance the robustness and efficiency of learning algorithms across diverse applications, from robotic control and video understanding to recommendation systems.