Deepak Pathak

Pathak earned a computer-science-and-engineering bachelor’s degree at IIT Kanpur and a PhD at UC Berkeley. He then spent a year as a researcher at Facebook AI Research and a visiting postdoc at Berkeley before joining Carnegie Mellon in 2020. He cofounded Skild AI in 2023 to develop a foundation model for robotics. His academic group studies how agents can acquire skills and generalize from raw sensory experience.[1][2]

Curiosity as a training signal

With Pulkit Agrawal, Alexei Efros and Trevor Darrell, Pathak developed an intrinsic curiosity module. It rewards an agent when the consequences of its actions are difficult to predict in a learned feature space. An inverse-dynamics objective helps focus that representation on aspects the agent can influence, rather than requiring accurate prediction of every pixel. Experiments in games examine exploration with sparse rewards, no external rewards and transfer to unfamiliar scenarios.[3]

Adapting a walking policy in real time

Pathak coauthored Rapid Motor Adaptation with Ashish Kumar, Zipeng Fu and Jitendra Malik. It pairs a base locomotion policy with an adaptation module that infers changing conditions from recent experience. Both components are trained in simulation, then deployed on a quadruped without additional fine-tuning. Tests include difficult surfaces, changing loads and disturbances. The architecture separates learning a useful movement policy from quickly estimating the conditions under which it must operate.[4]

In his Skild AI conference talk, Pathak argues for training across robot embodiments rather than creating a separate intelligence system for each machine. Simulation and human video provide pretraining material, while robot interaction supplies more direct experience. The ambition is a model whose general knowledge can transfer between physical forms and tasks, then improve through deployment.[5]