Deepak Pathak

Deepak Pathak’s research treats robotics as a problem of learning, not hand-crafted control. Across locomotion, manipulation and physical reasoning, he develops systems that extract supervision from experience, video and simulation—then tests whether one policy can adapt across bodies, tasks and environments.

Deepak Pathak is the Raj Reddy Associate Professor at Carnegie Mellon University and co-founder and chief executive of Skild AI. His academic work has long asked how agents can learn from curiosity and raw sensory experience rather than dense human instruction. That premise now informs a broad programme in generalist robotics. LocoFormer, a 2025 Conference on Robot Learning best-paper finalist co-authored by Pathak, trained on procedurally generated bodies and used long context to adapt one locomotion policy to unseen legged and wheeled robots, including changes in weight and motor failures. ViPRA, published for ICLR 2026, learned latent actions from unlabelled human and robot video; its authors reported gains of 16% on SIMPLER and 13% across real-world manipulation tasks against their selected baselines. Pathak also co-authored an ICLR 2026 oral paper on latent-particle world models, which learns object-centred scene dynamics from video without manual labels. In separate 2026 work, his team used physics simulators to generate reinforcement-learning supervision and reported five- to ten-point zero-shot gains on International Physics Olympiad problems. The studies are collaborative and research-scale. Their shared ambition is precise: turn diverse experience into reusable physical intelligence.