Uploaded October 2025 | Updated September 2026, 2 hours ago
Biological intelligence achieves remarkable generalization and rapid adaptation through compact, efficient representations. Inspired by insights from neuroscience, Sumeet's work seeks to give embodied artificial agents
similarly powerful abstractions for both behavior and perception. In this talk, he presents work in two main directions: (1) defining representations for behavior that enhance learning from expert data, and (2) developing self-supervised, structured visual representations from RGB data that enable zero-shot generalization for real-world robots. He concludes by outlining future research directions and discussing how deliberate, principled choices of representation can unlock greater generalizability in modern robotic systems.
Sumeet is a fifth-year PhD candidate at the University of Southern California, specializing in machine learning and robotics. His research draws inspiration from neuroscience to design representations that enable generalization and rapid adaptation in embodied agents. His work spans reinforcement learning for robotics, diffusion models as control policies, and self-supervised learning for extracting structured visual representations that support zero-shot generalization on real robots. During his internships with NVIDIA’s autonomous vehicles research team, he developed diffusion-based scenario generation tools and high-throughput simulation pipelines for RL training. His overarching goal is to create adaptable, robust, and intelligent robotic systems capable of operating effectively in complex, real-world environments.
Biological intelligence achieves remarkable generalization and rapid adaptation through compact, efficient representations. Inspired by insights from neuroscience, Sumeet's work seeks to give embodied artificial agents
similarly powerful abstractions for both behavior and perception. In this talk, he presents work in two main directions: (1) defining representations for behavior that enhance learning from expert data, and (2) developing self-supervised, structured visual representations from RGB data that enable zero-shot generalization for real-world robots. He concludes by outlining future research directions and discussing how deliberate, principled choices of representation can unlock greater generalizability in modern robotic systems.
Sumeet is a fifth-year PhD candidate at the University of Southern California, specializing in machine learning and robotics. His research draws inspiration from neuroscience to design representations that enable generalization and rapid adaptation in embodied agents. His work spans reinforcement learning for robotics, diffusion models as control policies, and self-supervised learning for extracting structured visual representations that support zero-shot generalization on real robots. During his internships with NVIDIA’s autonomous vehicles research team, he developed diffusion-based scenario generation tools and high-throughput simulation pipelines for RL training. His overarching goal is to create adaptable, robust, and intelligent robotic systems capable of operating effectively in complex, real-world environments.
