Deepak Pathak
Last reviewed
May 31, 2026
Sources
20 citations
Review status
Source-backed
Revision
v1 · 1,769 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 31, 2026
Sources
20 citations
Review status
Source-backed
Revision
v1 · 1,769 words
Add missing citations, update stale details, or suggest a clearer explanation.
Deepak Pathak is an Indian American roboticist and machine learning researcher who is the co-founder and chief executive officer of Skild AI, a startup building a general purpose foundation model for robotics. He is also a faculty member in the School of Computer Science at Carnegie Mellon University, where he holds the Raj Reddy associate professorship with appointments in the Robotics Institute and the Machine Learning Department. His research spans reinforcement learning, computer vision, and robot learning, with influential contributions to curiosity driven exploration, self-supervised learning, and methods that let robots adapt and generalize across different bodies and environments.[1][2]
Pathak is known in academic circles for the 2017 paper on curiosity driven exploration through self-supervised prediction and the 2016 paper on context encoders, both of which became widely cited reference points in their subfields. In 2023 he left full time academia, with his Carnegie Mellon colleague Abhinav Gupta, to start Skild AI, which by early 2026 had raised more than two billion dollars and reached a valuation above fourteen billion dollars.[3][4]
Pathak grew up in India and studied computer science at the Indian Institute of Technology Kanpur, where he earned a bachelor's degree and graduated as a gold medalist for top academic performance in his department.[2][5] He later received the Young Alumnus Award from IIT Kanpur in 2022.[1]
He moved to the United States for graduate study at the University of California, Berkeley, completing his PhD in 2019. His doctoral advisors were Alexei Efros and Trevor Darrell, two researchers in computer vision and machine learning at Berkeley.[1][6] During his time at Berkeley he co-founded VisageMap, a facial recognition startup that was later acquired by FaceFirst.[7] After finishing his doctorate he spent a period as a visiting postdoctoral researcher at Berkeley working with Pieter Abbeel before joining the faculty at Carnegie Mellon.[1]
Pathak's research sits at the intersection of computer vision, machine learning, and robotics, with a recurring focus on how agents can learn general behaviors from raw sensory input without heavy human supervision.[1]
One of Pathak's most cited works is the 2017 paper "Curiosity-driven Exploration by Self-supervised Prediction," published at the International Conference on Machine Learning. The paper formulates curiosity as the error in an agent's ability to predict the consequence of its own actions in a visual feature space that is itself learned through a self-supervised inverse dynamics model. This intrinsic reward signal lets an agent explore and learn even when external rewards are sparse or absent. The method was evaluated in game environments including VizDoom and Super Mario Bros, and it became a standard reference for intrinsically motivated reinforcement learning.[8][9]
Pathak extended this line of work with a follow up approach that frames exploration as disagreement. Rather than relying on a single predictive model, an ensemble of dynamics models is trained, and the agent seeks out states where the models disagree, which provides a differentiable exploration signal that works in stochastic settings.[10]
While at Berkeley, Pathak was the lead author of "Context Encoders: Feature Learning by Inpainting," presented at the 2016 Conference on Computer Vision and Pattern Recognition. The paper introduced a convolutional network trained to generate the contents of a missing image region conditioned on its surroundings, an image inpainting task. To fill in the gap, the network must understand both the appearance and the semantics of the scene, so the features it learns transfer to downstream tasks such as classification, detection, and segmentation. The work combined a pixel reconstruction loss with an adversarial loss and became an early example of self-supervised representation learning for vision.[11][12]
A central theme of Pathak's robotics research is building controllers that generalize beyond a single robot or a single environment. With collaborators at Carnegie Mellon and Berkeley, including Ashish Kumar, Zipeng Fu, and Jitendra Malik, he developed Rapid Motor Adaptation, or RMA, an algorithm that lets a legged robot adapt online to new terrain and conditions within a fraction of a second. RMA pairs a base policy trained with reinforcement learning in simulation with an adaptation module that infers properties of the environment from the robot's own movements. It was trained entirely in simulation and deployed on a quadruped robot without fine tuning, and because the approach is learning based it is not tied to a single robot design.[13][14]
This interest in cross embodiment generalization, the idea that one learned model can drive many different robot bodies, later became a guiding principle for Skild AI. Between his Berkeley doctorate and his startup, Pathak also spent time as a researcher at Meta AI Research, where he collaborated with Jitendra Malik.[1]
Pathak co-founded Skild AI in 2023 with Abhinav Gupta, who had been a colleague at Carnegie Mellon. Both founders left their academic teaching roles to build the company, with Pathak serving as chief executive officer and Gupta as a co-founder.[3][7] The company is headquartered in Pittsburgh, the same city as Carnegie Mellon.[15]
Skild AI develops what it calls the Skild Brain, described as a unified, omni-bodied model meant to control many kinds of robot for many kinds of task. The system is built to perform low level skills such as grasping, handover, locomotion, and navigation, which are then exposed through an interface so that applications can be built on top of them. The company positions learning from human video as a way to address the shortage of robot training data, and it has shown the model running on platforms aimed at security and inspection work, mobile manipulation, and automated packing.[16][2]
The approach reflects the foundation model paradigm that reshaped language and vision applied to physical machines. Rather than training a separate controller for each robot and task, Skild AI aims for a single model that transfers across hardware, an extension of the cross embodiment generalization Pathak studied in academia.[3]
Skild AI drew large investments soon after it emerged from stealth. In July 2024 the company announced a three hundred million dollar Series A round that valued it at about one and a half billion dollars. The round was led by Lightspeed Venture Partners, Coatue, SoftBank Group, and Jeff Bezos through Bezos Expeditions, with participation from Felicis, Sequoia, Menlo Ventures, General Catalyst, CRV, SV Angel, Carnegie Mellon University, and Amazon funds.[17][18]
In mid 2025 the company raised a further round led by SoftBank, reported at roughly five hundred million dollars, that lifted its valuation to about four and a half billion dollars, with Nvidia and Samsung among the new backers.[19] In January 2026 Skild AI announced a one and a half billion dollar scale round, reported as a 1.4 billion dollar raise led by SoftBank, that valued the company above fourteen billion dollars and included Nvidia, Macquarie Group, and 1789 Capital among others. By that point Pathak said the company had raised more than two billion dollars in total.[3][4]
The scale of this funding placed Skild AI among the most heavily capitalized robotics software startups, and it reflected wider investor interest in physical AI and the prospect of general purpose robot intelligence. Pathak has framed the long term goal as artificial general intelligence grounded in the physical world.[2][20]
At Carnegie Mellon, Pathak joined the School of Computer Science as an assistant professor and was later named to the Raj Reddy associate professorship in 2024, with affiliations in the Robotics Institute and the Machine Learning Department.[1] He has supervised graduate students and led a research group focused on agents that learn to act in the real world, and his lab's work on legged locomotion and self-supervised learning has been featured in the university's research news.[14] He continued to hold a faculty appointment at the university while leading Skild AI.[1]
Pathak has received a number of awards across his research career. He was named to the MIT Technology Review list of innovators under 35 in 2024 and received a Sloan Research Fellowship in 2025. Earlier honors include the Okawa Research Award in 2022, a Google Faculty Research Award in 2020, and graduate fellowships from Facebook, Nvidia, and Snap during his doctoral years. His CMU profile also lists an Office of Naval Research Young Investigator Award.[1]
| Field | Detail |
|---|---|
| Full name | Deepak Pathak |
| Known for | Curiosity driven exploration, context encoders, robot learning, Skild AI |
| Fields | Robotics, machine learning, computer vision |
| Current roles | Co-founder and CEO of Skild AI; faculty at Carnegie Mellon University |
| Education | BTech, IIT Kanpur; PhD, UC Berkeley (2019) |
| Doctoral advisors | Alexei Efros and Trevor Darrell |
| Skild AI founded | 2023 (with Abhinav Gupta) |
| Skild AI valuation | Above 14 billion dollars (January 2026) |
| Notable awards | MIT TR35 (2024), Sloan Research Fellowship (2025), Okawa Award (2022) |