Hanchen Cui

I am a Ph.D. student in computer science at the University of Minnesota Twin Cities, working with Prof. Karthik Desingh. I received my master's and bachelor's degrees from Shanghai Jiao Tong University and Xidian University, respectively. Previously, I was a researcher at Shanghai Qi Zhi Institute, where I worked closely with Prof. Yang Gao at IIIS, Tsinghua University.

My research interest lies in robot learning and my long-term goal is to build a generalized robot that can accomplish long-horizon and complex tasks. To be more specific, I work on world models, vision-language-action models, and multimodal representation learning.

In summer 2025, I interned at Meta FAIR, where I was mentored by Franziska Meier and Krishna Murthy Jatavallabhula and worked on action-conditioned world models and vision-language-action models. In summer 2026, I interned at Amazon Robotics (Fauna Robotics team), where I worked with Roberto Martin-Martin and Luca Carlone. During the internship, I worked on (1) sim-real co-training for VLA models, and (2) latent world-action models.

Email: cui00191[at]umn.edu

Email / Scholar / LinkedIn

Publications

* indicates equal contribution and † indicates joint last author. Representative work is highlighted.

	PILOT: Guiding Long-Horizon World-Model Planning for Robot Manipulation with Visual Subgoals and Policy Priors Yang Liu, Hanchen Cui, Karthik Desingh project page / arXiv We propose PILOT, which uses video-generated visual subgoals to guide long-horizon robot manipulation and policy priors to initialize world-model-based action optimization. PILOT achieves 68.8% success on real-world manipulation tasks, outperforming direct policy execution and improving planning efficiency.
	Active Real-World Factor-Based Evaluation for Generalist Robot Policies Andrew Liao, Hanchen Cui, Karthik Desingh, Aryan Deshwal arXiv We propose a Bayesian active evaluation framework that efficiently assesses generalist robot policies by adaptively selecting informative real-world test configurations. Across three manipulation tasks, our approach reduces evaluation trials by 20–40% compared to random testing while accurately characterizing policy generalization and identifying failure-prone conditions.
	DREAMSTEER: Latent World Models can Steer VLA Policies during Deployment Without Any Finetuning Hanchen Cui, Sergio Arnaud, Arjun Majumdar, Daniel Dugas, Elie Aljalbout, Karthik Desingh^†, Krishna Murthy Jatavallabhula^† Franziska Meier^† project page / arXiv We propose DREAMSTEER, a plug-and-play deployment-time steering framework that enhances pretrained VLA policies without any finetuning, or any in-domain data. By imagining candidate futures with a latent world model and ranking them with a value model, DREAMSTEER enables robots to select more reliable action sequences before execution.
	Fine-Tuning Hard-to-Simulate Objectives for Quadruped Locomotion: A Case Study on Total Power Saving Ruiqian Nai, Jiacheng You, Liu Cao, Hanchen Cui, Shiyuan Zhang, Huazhe Xu, Yang Gao ICRA 2025 project page / arXiv We propose a data-driven framework for fine-tuning locomotion policies, targeting these hard-to-simulate objectives. Our framework leverages real-world data to model these objectives and incorporates the learned model into simulation for policy improvement.
	A Universal World Model Learned from Large Scale and Diverse Videos Hanchen Cui, Yang Gao NeurIPS 2023, Foundation Models for Decision Making Workshop pdf / poster / code We propose a generalizable world model pre-trained by large scale and diverse video dataset using latent action extracted by VQ-VAE and then fine-tune the world model by robot data to obtain an accurate dynamic function.
	A Universal Semantic-Geometric Representation for Robotic Manipulation Tong Zhang, Yingdong Hu, Hanchen Cui,Hang Zhao, Yang Gao CoRL 2023 project page / arXiv / code We present Semantic-Geometric Representation (SGR), a universal perception module for robotics that leverages the rich semantic information of large-scale pre-trained 2D models and inherits the merits of 3D spatial reasoning.

Research projects

	Sim-Real Co-Training for VLA models (1) Built a sim-real co-training pipeline for Unitree G1 manipulation using 3,000 simulation trajectories and real-world demonstrations. (2) Studied sim-to-real transfer and generalization by varying co-training ratio, simulation data scale, and domain randomization. (3) Proposed real-image-based simulation augmentation, improving real-world success rate from 31.8% to 45% in unseen scenes with distractors.
	A real-time vision-language manipulation system (1) Build a vision-language manipulation deployment system to accomplish a variety of real world tasks based on Franka Panda Robot. (2) A robust client-server architecture is employed to enable low-latency remote control. (3) Fine-tune a vision-language manipulation model (CLIPORT), achieving 90% success rate in real-world experiments.
	Legged robots learn from physical world (1) Establish a sim2real locomotion deployment codebase for legged robots. (2) Develop a real-world learning framework that acquires data from real-world interactions and performs policy training on a remote server.
	Visual planning for legged robots (1) Leverage the visual understanding and high-level planning abilities of large multimodal models (GPT4o), enabling legged robots to accomplish multi-step and complex tasks, such as sending packages, and objects retrieval.

Services

Reviewers: CoRL, T-RL, ICRA, RA-L, IROS

Modified version of template from here.