Mila > Équipe > Siddarth Venkatraman

Siddarth Venkatraman

I am a PhD student working with Prof. Glen Berseth at the Robotics and Embodied AI Lab (REAL). My research revolves around figuring out interesting ways to leverage generative models in reinforcement learning. I am particularly interested in unsupervised discovery of skill representations that can be utilized as strong priors for downstream policies optimized for novel objectives. I believe diverse skill discovery requires expressive, multimodal policy distributions which are hard to train in the context of reinforcement learning. I expect the path to solving this problem involves developing techniques that facilitate training expressive energy based policies in an efficient and scalable manner. Diffusion generative models are a powerful framework to help achieve this goal, and I am currently working on various projects which involve learning diffusion based policies.