JEPA-Style World Models with a Learned Critic for Long-Horizon Robot Planning
Coming SoonCS PhD Student · Oklahoma State University
Chrisantus Eze
I'm a Computer Science PhD student at Oklahoma State University, advised by Christopher Crick. I build systems that enable robots to reason about and manipulate objects in complex, cluttered environments — combining foundation models (VLMs, VLAs), imitation learning, and model-based planning to support long-horizon spatial reasoning.
I'm a fourth-year PhD student at Oklahoma State University, previously a Graduate Student Mentee at Google (Feb 2023 – Jun 2025), where I co-developed sequential manipulation policies achieving 98% benchmark success. I received my B.Eng. in Electrical & Electronic Engineering from FUTO, Nigeria.
My current projects include Unveiler — a planning–control framework for sequential object retrieval in clutter (94% sim-to-real success) — and VLM-RAG pipelines for object grounding (+34% improvement). I also developed A3, a source-free domain adaptation method that improves cross-domain transfer by 15%. I am currently working on a JEPA-style world model with a learned critic for long-horizon robot planning, and exploring logic-augmented spatial reasoning via Answer Set Programming (ASP) to improve compositional reasoning in VLMs. I have published 6 peer-reviewed papers across robotics and ML venues.
chrisantus.eze at okstate dot edu
