Bio

Jing Zhang is a Postdoctoral Scholar at NYU, jointly appointed in the Center for Robotics and Embodied Intelligence and the Department of Anthropology. She is advised by Chen Feng (AI4CE Lab) and Radu Iovita (Anthrotopography Lab). She received her Ph.D. in photogrammetry from Wuhan University. Over the next three years, she has two goals: (1) build agents that autonomously navigate New York City, use tools, and assist people safely; and (2) assemble and rectify fossil fragments to reveal patterns of human evolution. She advances these aims through Evolving Embodied Intelligence, a cognitively inspired framework that integrates perception, imagination, reasoning, action, and feedback in a closed loop across physical and digital worlds. Her interdisciplinary work has appeared in top venues in computer vision (CVPR, ICCV) and robotics (ICRA), and in leading journals spanning engineering and optics (IEEE Transactions on Industrial Informatics, Photonics Research, Optics Express, Optics Letters).

Evolving Emboided Intelligence via Perception, Imagination, and Action Loops

I found myself at a unique multidisciplinary intersection: collaborating with AI and robotics researchers building autonomous systems for the future, and anthropologists reconstructing fragments of the human past. Working across these two worlds, I began to see a deeper parallel: whether navigating a robot dog through the crowded streets of New York City or interpreting the mystery of a deformed fossil skull from a cave in South Africa, intelligent systems must operate under uncertainty, reason over partial information, and iteratively refine their action. This realization led me to a unifying question: How can we build agents that learn not just by seeing, but by simulating, acting, and evolving through feedback?

This question forms the foundation of my research framework:Evolving Embodied Agents via Perception, Imagination, and Action Loops. Within this framework, the core challenge is not to solve perception, imagination, and action in isolation, but to couple them into a self-evolving cycle through which agents refine their intelligence over time: (1) How can agents perceive meaningfully from incomplete and noisy observations? Instead of passively encoding raw geometry, perception must extract task-driven structure—whether from limited-view sensor data or fragmented artifacts. (2) How can agents imagine what is missing or yet to come? When observations are partial, agents must infer hidden structures and predict plausible futures, grounding imagination in perception while leveraging world models and commonsense priors to propose testable hypotheses. (3) How can agents act to probe the world and adapt? Actions are not only plan executions but interventions that alter the environment, generating new evidence that updates perception and reshapes imagination, thereby closing the loop for continual task-driven learning.