Hi there! Iβm Gokul, a final-year PhD candidate in the Robotics Institute at Carnegie Mellon University, working on efficient algorithms for interactive learning (e.g., imitation / RL / RLHF).
My research focuses on developing the novel algorithmic paradigms required to build robustly aligned agents that gracefully handle situations unseen in their training data. In other words, rather than giving agents the βfishβ, my research focuses on teaching agents to fish. I fuse ideas from RL and game theory to develop principled and scalable algorithms for domains like robotic manipulation and language modeling.
I work with Drew Bagnell and Steven Wu. I completed my B.S. / M.S. at UC Berkeley, where I worked with Anca Dragan on Learning with Humans in the Loop. Iβve spent summers working on ML @ SpaceX, Perception @ NVIDIA, Motion Planning @ Aurora, World Models @ Microsoft and LLMs @ Google.
π I am currently on the job market! π
Events & News
June 2025 - Two new papers out on learning to search: SAILOR that outperforms diffusion policies trained on 10x as many demos on multi-stage visual manipulation tasks (Spotlight @ NeurIPS β25), FOREWARN that allows real robots to avoid complex semantic failures via VLM verifiers (published at RSS β25).
March 2025 - New particularly exciting preprint out on the real value of RL in fine-tuning / RLHF. I gave a talk at Cornell on the paper that might also be of interest.
November 2024 - Drew, Steven, and I are co-teaching a course on the algorithmic foundations of interactive learning. If youβd like to understand the fundamental principles behind imitation (e.g. for robots) and RLHF (e.g. for LLMs), this is the course for you!
Research Highlights
A Smooth Sea Never Made a Skilled ππ°πΈπ»πΎπ: Robust Imitation via Learning to Search
We introduce ππ°πΈπ»πΎπ: a method for learning to search from expert demonstrations that out-performs Diffusion Policies trained in 5-10x as much data on multi-stage visual manipulation tasks. [Paper] [Site] [Podcast]
All Roads Lead to Likelihood
We explore how the value of RL in fine-tuning / RLHF seems to be fundamentally derived from generation-verification gaps. [Paper] [Talk]
SPO: Self-Play Preference Optimization
We derive a new fundamental algorithm for RLHF that robustly handles the complex, intransitive preferences that often result from aggregating a diversity of views. [Website] [Paper]
Inverse RL without RL
We derive exponentially faster algorithms for inverse RL by proving that local search is "all you need" for imitation. [Website] [Paper] [Podcast]
Of Moments and Matching
We provide a unifying, game-theoretic framework for imitation learning that explains when different algorithmic families can avoid compounding errors. [Website] [Blog]