PWM: Policy Learning with Large World Models
Ignat Georgiev, Varun Giridhar, Nicklas Hansen, Animesh Garg
Hey! I'm Ignat, a Robot Learning researcher. I am currently doing a PhD at Georgia Tech, working with Animesh Garg. My deeply seeded belief is that robotics and learning have the potential to change the world for the better, and my mission is to bring intelligent star-wars-like robots one step closer to reality (minus the blasters).
Prior used to work as an ML engineer/researcher in autonomous vehicles. Nowadays I occasionally do some consulting or contracting work on projects I find interesting. Shoot me an email if you have something in mind!
This little website is my space to tell the world about me, put my resume out there, show off my research and blog some wisdom. You can find all of my research on Google Scholar.
Selected Research
Actor-Critic for Policy Learning in Contact-Rich Differentiable Simulation
Ignat Georgiev, Krishnan Srinivasan, Jie Xu, Eric Heiden, Animesh Garg
Iterative Semi-parametric Dynamics Model Learning For Autonomous Racing
Ignat Georgiev, Christoforos Chatzikomis, Timo Völkl, Joshua Smith, Michael Mistry
Blog posts
Why Behaviour Cloning?
How BC solved tasks plaguing RL for years
Everybody wants to love Reinforcement Learning (RL)! It was the 3rd most popular topic at NeurIPS 2024, just after CV and NLP 🤯! RL has this elegant mathematical framework that mirrors how we humans learn through trial and error. I even identify as an RL person! But in the realm...
[Read More]
What makes RL tick?
Exploring one of the less appreciated parts of RL
For decades, humans have been fascinated with controlling complex robots to do human-like dexterous skills. Building upon decades of control theory research, Boston Dynamics can make Atlas do some pretty cool things that only very few humans can do!
[Read More]
Model Based Reinforcement Learning
An overview of what people have been up to
When most people hear Reinforcement Learning (RL), they usually think about some of the key works in the field, such as Deepmind’s DQN playing Atari, OpenAI’s robot hand solving Rubik’s cube or more recently, Reinforcement Learning from Human Feedback (RLHF) that powers ChatGPT. All of these assume that they are...
[Read More]
Deriving Soft Actor Critic (SAC)
Going from zero to one of the key RL algorithms - Soft Actor Critic (SAC)
Soft Actor Critic (SAC) is a model-free Reinforcement Learning (RL) method developed by Tuomas Haarnoja in a series of papers culminating in Soft Actor-Critic Algorithms and Applications. It is one of the seminal RL papers for continuous control and has inspired much of my work. In this blog post, I’ll...
[Read More]
Deriving Differential Dynamic Programming
Revisiting a classic trajectory optimization approach
Differential Dynamic Programming (DDP), first proposed by David Mayne in 1965 is one of the oldest trajectory optimization techniques in optimal control literature. It is an extension of Dynamic Programming where instead of optimising over the full state space, we are only optimise around a nominal trajectory by taking 2nd...
[Read More]