Deepak Pathak

Deepak Pathak is the co-founder and CEO of Skild AI, building a general-purpose robot brain that can learn across different machines, tasks, and environments. His research connects self-supervised perception, curiosity-driven exploration, and adaptive robot control: ways for machines to learn from images, their own actions, and changing physical conditions without a person specifying every lesson.

From learning to see to learning to act

Pathak earned his computer-science bachelor’s degree at IIT Kanpur in 2014 and completed his PhD at UC Berkeley in 2019, advised by Alexei Efros and Trevor Darrell. During his doctoral years, he also co-founded VisageMap, which was later acquired by FaceFirst. He then spent a year as a researcher at Facebook AI Research working with Jitendra Malik and as a visiting postdoctoral researcher at Berkeley with Pieter Abbeel, before joining Carnegie Mellon in 2020. He is the university’s Raj Reddy Associate Professor in the Robotics Institute, affiliated with the Machine Learning Department.

An early contribution was Context Encoders, developed with collaborators in 2016. The model learned visual representations by reconstructing missing image regions. Filling a hole required understanding the surrounding scene well enough to generate a plausible replacement; the image itself supplied the training target. Prediction became a way to learn useful visual features without manually labeling every example.

Pathak carried that approach into action. His collaborative 2017 research on curiosity-driven exploration gave an agent an internal reward when it failed to predict the consequences of its own behavior. The reward operated in a learned visual representation rather than directly on pixels, helping the agent concentrate on aspects of the environment relevant to its actions. Experiments in Super Mario Bros. and VizDoom explored how agents could discover useful behavior with sparse external rewards—or none at all.

At Carnegie Mellon, his work increasingly addressed how learned behavior could adapt to the physical world. Two complementary directions emerged: adjusting control to changing conditions and extracting useful guidance from human activity.

Learning beyond a fixed training setup

• Rapid Motor Adaptation: With Ashish Kumar, Zipeng Fu, and Jitendra Malik, Pathak developed RMA in 2021. A base control policy and an adaptation module, both trained in simulation, enabled a quadruped to adjust to unfamiliar terrain and changing conditions in fractions of a second. The system transferred to a physical robot without fine-tuning, addressing the practical problem that ground surfaces, loads, and hardware conditions rarely stay as they were during training.

• Learning from human video: Pathak’s collaborative work on WHIRL used a human demonstration to initialize robot learning, followed by physical interaction to improve the behavior. The subsequent Vision-Robotics Bridge, developed with Shikhar Bahl, Russell Mendonca, Lili Chen, and Unnat Jain, learned where and how people interact with objects from large human-video datasets. For a drawer, that means identifying the handle as the contact point and an outward movement as the useful trajectory. These affordances give a robot a more directed starting point for learning, even when its surroundings differ from the human’s.

• Omni-bodied intelligence: At Skild, Pathak pursues a shared model across robot bodies. His approach to robotics data weighs scalability, diversity, and proximity to the robot’s own action measurements. Simulation scales readily but requires scenes to be engineered; human video supplies varied situations but requires mapping human behavior to robot actions; teleoperation supplies precise robot examples in smaller quantities. Skild combines simulation and human video for pre-training with teleoperation for post-training. Learning across hardware allows those sources to contribute to a shared model rather than a separate learning system for every machine.

Building Skild around deployment

Pathak founded Skild AI with Abhinav Gupta in 2023; the company emerged from stealth in July 2024. Its development extends his academic work toward a model that can be adapted to industrial and everyday tasks.

Skild’s S1 model, introduced in August 2026, uses video demonstrations as prompts for robot behavior. Its training pairs demonstrations with action episodes, encouraging the model to infer intent across differences in scene, viewpoint, and embodiment. The aim is to teach a new task through context without updating the model’s weights. Skild demonstrated this approach on plant potting, coffee preparation, and kit assembly.

Pathak and Gupta argue that deployment is part of research. Their writing on production robotics treats installation, maintenance, changing processes, and cycle time as questions that must shape the technology from the beginning. Their egg-cooking example makes the cost of reliability concrete: achieving the first success took a week, while making it reliable across different eggs and setups took another two months. On a factory line, accurate execution also has to be fast enough to keep pace with neighboring stations.

That experience feeds their deployment data flywheel: specialize a general model for different applications, then bring experience from those applications back into the shared model. For Pathak, generality has a practical purpose. Learning from one robot should reduce the work needed to make the next robot useful.