Liu Cao

Email / Scholar / Github

Hello there! I'm Liu Cao, currently a first-year PhD student in IIIS, Tsinghua University, advised by Prof. Mengdi Xu on Humanoid Robots and Whole-body Control. Before that, I received my bachelor's degree from EE, Tsinghua University.

My current interests lie in developing autonomous robots, particularly humanoid and mobile manipulators, capable of active perception and interacting in the real-world scenarios.

I am lucky to work closely with Botian Xu.

Research

WEAVE: Learning Whole-Body Dexterous Loco-Manipulation from Human–Object Interactions
Liu Cao, Xingze Wu, Jingzhi Cui, Botian Xu, Mingzhi Pei, Ruoqu Chen, Mengdi Xu
arXiv preprint, 2026
project page / pdf / code

We present Weave, a framework for learning whole-body dexterous humanoid–object interaction from captured human demonstrations. It reaches 92.5% success on trained interactions and 65.0% on unseen sequences, and we release ~23 hours of physically executed rollouts.

HuMiT: Low-Latency Whole-Body Humanoid Teleoperation via Minimal Reference Tracking
Liu Cao, Botian Xu, Ruoqu Chen, Jiajun Liu, Mingzhi Pei, Zihao Wang, Zhongyu Li, Mengdi Xu
Under Review, 2026
project page

We present HuMiT, a whole-body teleoperation system that drives a humanoid from sparse human keypoints alone, achieving ~20 ms communication-to-inference latency to enable reactive whole-body skills such as ball catching and loco-manipulation.

DEX-X: Learning Visual-Tactile Dexterous Manipulation From Human Videos with Simulated Interaction
Ruoqu Chen, Feixiang Ruan, Liu Cao, Zihao Wang, Botian Xu, Shiqin Tong, Jiajun Liu, Mingzhi Pei, Chenyu Zhang, Wanli Xing, Kaifeng Zhang, Mengdi Xu
Conference on Robot Learning (CoRL), 2026
project page / arXiv

DEX-X learns visual-tactile dexterous manipulation from monocular human videos by replaying them in simulation to recover the missing contact signal. The distilled policies transfer zero-shot to a real hand-arm platform, reaching 93% success on cube picking.

What Do VLAs Actually Learn through In-Context Failure Conditioning?
Jiajun Liu, Jieming Li, Zi Zhuang, Hang Yu, Qingli Chen, Liu Cao, Yingxi Lu, Ruoqu Chen, Yuhang Cao, Chenyu Zhang, Yankai Lin, Mengdi Xu
3D-LLM/VLA Workshop, CVPR, 2026
project page / code

Using RoboRetry as a controlled probe across 12 RLBench tasks, we study whether vision-language-action policies actually learn from prior failures via in-context conditioning, disentangling a slot-presence effect from content-dependent gains, and release FailureSlot, a dataset of 1,334 annotated failure trajectories.

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,
IEEE International Conference on Robotics and Automation (ICRA), 2025
project page / arXiv

Better locomotion using real-world data. Our approach achieves a 24-28% net reduction in power consumption.

Hybrid Internal Model: Learning Agile Legged Locomotion with Simulated Robot Response
Junfeng Long^*, Zirui Wang^*, Quanyi Li, Jiawei Gao, Liu Cao, Jiangmiao Pang
^*Equal contribution
International Conference on Learning Representations (ICLR), 2024
project page / arXiv

We present the Hybrid Internal Model, a method enabling the control policy to estimate environmental disturbances by only explicitly estimating velocity and implicitly simulating the system's response.

Detecting Vulnerable Nodes — teaser figure

Detecting Vulnerable Nodes in Urban Infrastructure Interdependent Network
Jinzhu Mao^*, Liu Cao^*, Chen Gao, Huandong Wang, Hangyu Fan, Depeng Jin, Yong Li
^*Equal contribution
ACM SIGKDD Conference on Knowledge Discovery and Data Mining (SIGKDD), 2023
code / arXiv

We model the interdependent network as a heterogeneous graph and propose a system based on graph neural network with reinforcement learning, which can be trained on real-world data, to characterize the vulnerability of the city system accurately.