Ajay Sridhar

I am a Computer Science PhD student at Stanford University, co-advised by Dorsa Sadigh and Chelsea Finn. My research is supported by the NSF Graduate Research Fellowship. I am interested in building generalizable and robust robot learning systems that continuously improve with experience.

Previously, I was an undergraduate at UC Berkeley, where I worked with Sergey Levine and Dhruv Shah at the Robotic AI and Learning Lab. I have also worked with Thomas Dietterich.

Email  /  Google Scholar  /  Github  /  CV  /  Twitter

MemER is a hierarchical policy framework, where the high-level policy is trained to select and track previous task-relevant keyframes from its experience. The high-level policy uses selected keyframes and the most recent frames when generating text instructions for a low-level policy to execute.

LeLaN is a language-conditioned navigation policy that learns from in-the-wild videos to enable natural language navigation commands for mobile robots.

SELFI is an online reinforcement learning approach for fine-tuning control policies trained with model-based learning. We combine the objective used during model-based learning with a Q-value function learned online.

NoMaD is a novel architecture for robotic navigation in previously unseen environments that uses a unified diffusion policy to jointly represent exploratory task-agnostic behavior and goal-directed task-specific behavior.

ViNT is a flexible Transformer-based model for visual navigation that can be efficiently adapated to a variety of downstream navigational tasks.

SACSoN is vision-based navigation policy that learns socially unobtrusive behavior in human-occupied spaces through continual learning.

GNM is vision-based navigation policy trained with a simple goal-reaching objective on a cross-embodiment navigation dataset. It exhibits positive transfer, outperforming specialist models trained on singular embodiment datasets, and generalizes to new robots.

ExAug is a vision-based navigation policy that learns to control robots with varying camera types, camera placements, robot sizes, and velocity constraints by applying a novel geometric-aware objective to view augmented data.

Source code from Jon Barron's website.