Chaoqi Liu

I'm a 4th year undergraduate student at UIUC. I like robotics. I'm currently visiting Yilun Du's lab. I've had the opportunity to be advised by Kris Hauser, Yunzhu Li, Maxim Likhachev.

I build robots that act intelligently. I began with kinodynamic motion planning (IxG), which assumes a reliable model of the environment. To handle broader settings without accurate models, I moved to learned dynamics for planning— graph-based neural dynamics (GBND) and pixel-space video predictors (HMA). In practice, GBND required heavy hand-engineering for graph instantiation, and contemporary video models were not accurate or controllable enough for planning, so I pivoted to learning to act directly—focusing on modular policies and studying implicit factorization of the action distribution (FDP) and explicit factorization by sensory modality (MMPC). This line of work crystallized a view: raw action space is the wrong abstraction. Thus, my current research pursues a Matryoshka-style action tokenization that imposes causal structure on action, enables autoregressive reasoning over semantically meaningful actions.

Email | Google Scholar | Github | Twitter 𝕏

Research

	Haonan Chen, Jiaming Xu, Hongyu Chen, Kaiwen Hong, Binghao Huang, Chaoqi Liu, Jiayuan Mao, Yunzhu Li, Yilun Du⁺, Katherine Driggs-Campbell⁺ Multi-Modal Manipulation via Multi-Modal Policy Consensus 2025, [webpage], [arXiv], [code] Multi-Modal Policy Consensus (MMPC) decomposes control across diffusion experts specialized for vision, touch, and other modalities, with a router network that adaptively balances their influence—achieving flexible multimodal fusion.
	Chaoqi Liu, Haonan Chen, Sigmund H. Høeg, Shaoxiong Yao, Yunzhu Li, Kris Hauser, Yilun Du, Flexible Multitask Learning with Factorized Diffusion Policy 2025, [webpage], [paper] Spotlight at RSS 2025 2nd Workshop on Semantic Reasoning and Goal Understanding in Robotics [Link] Factorized Diffusion Policy (FDP) is a modular robot policy framework that composes specialized diffusion components through score aggregation, enabling scalable multitask learning, interpretable behavior decomposition, and efficient task adaptation.
	Sigmund H. Høeg, Aksel Vaaler, Chaoqi Liu, Olav Egeland, Yilun Du, Hybrid Diffusion for Simultaneous Symbolic and Continuous Planning 2025, [webpage], [arXiv] Hybrid Diffusion Policy (HDP) is a long-horizon robot planning framework that composes discrete high-level plans and continuous low-level trajectories through a unified diffusion process, enabling robust multimodal behavior generation, improved decision-making over long horizons, and flexible conditioning for complex tasks.
	Lirui Wang, Kevin Zhao, Chaoqi Liu, Xinlei Chen Learning Real-World Action-Video Dynamics with Heterogeneous Masked Autoregression 2025, [webpage], [arXiv] Heterogeneous Masked Autoregression (HMA) is a real-time robotic video simulation for high-fidelity and controllable interactions, leveraging masked autoregressive models and heterogeneous training.
	Chaoqi Liu, Yunzhu Li, Kris Hauser Localized Graph-Based Neural Dynamics Models for Terrain Manipulation 2025, [webpage], [arXiv] A scalable and efficient graph-based neural terrain dynamics (gbnd) by introducing fine-grained adaptive and automatic particle selection. The proposed framework is both orders of magnitude efficient and more accurate than previous gbnd.
	Ramkumar Natarajan, Chaoqi Liu, Howie Choset, Maxim Likhachev Implicit Graph Search for Planning on Graphs of Convex Sets RSS 2024, [webpage], [paper], [arXiv] A scalable and efficient way to plan on Graphs of Convex Sets (GCS) with stronger theoretical properties. We plan on GCS using a previously developed hybrid search-optimization framework called INSAT.

Misc.

I have consistently maintained a perfect 4.0 GPA, excelling with top grades of A/A+ in all of my Computer Science, Mathematics, and Statistics courses. A complete list can be found here.

Page template borrowed from Jon Barron