Built a spatially adaptive CNN that removes spatially varying optical blur by learning point spread functions across the image field, trained on a ray-traced synthetic pipeline with PSF-based augmentation; improved restoration quality by 1.7% over prior baselines.
I am a first-year Ph.D. student in Electrical Engineering and Computer Sciences at UC Berkeley. Before Berkeley, I received B.S. degrees from the University of Southern California in Electrical and Computer Engineering (Honors) and Applied & Computational Mathematics. At USC, I worked with Prof. Bhaskar Krishnamachari in the Autonomous Networks Research Group and with Prof. Peter Beerel. I have also worked with Prof. Yinyu Ye at Stanford, Prof. Qi Zhu as a visiting scholar in the Design Automation of Intelligent Systems Lab at Northwestern, and Prof. Dionysis Kalogerias at Yale.
Research Vision
My research aims to advance general-purpose physical intelligence by developing robots and multi-agent systems that learn to perceive, plan, and act reliably under uncertainty, with meaningful guarantees on their behavior.
To this end, I study predictive world models for reasoning under partial observability, risk-aware decision-making and multi-agent coordination, and specification-guided learning and verification for robotic policies, including vision-language-action models. My goal is to develop adaptable robots that generalize across tasks and environments, improve through experience, and remain safe and robust with clearly defined guarantees.
Education
B.S. in Applied & Computational Mathematics
University of Southern California
News
- Sep 2026SENTINEL is accepted to NeurIPS 2026! 🎉
- Aug 2026Started my Ph.D. in EECS at UC Berkeley.
- Dec 2025DISPATCH released on arXiv.
- Nov 2025Predictive Modeling in AUV Navigation submitted to IEEE IV 2026 (arXiv).
- Sep 2025SENTINEL released, with a project page.
- Jun 2025Distributed ML inference paper published at GLSVLSI 2025.
- Jun 2025Started as a visiting scholar at the Design Automation of Intelligent Systems Lab, Northwestern.
- May 2025Undergraduate thesis on online dynamic resource allocation accepted by USC.
- Nov 2024Pedestrian detection paper published at IEEE SiPS 2024.
- May 2024Honorable Mention, MCM 2024, for Kalman-filter-based submarine trajectory prediction.
- Jun 2023Started as an ML intern at Tsinghua MetaCam Lab.
Selected Publications
* Equal contribution · Full list
AnyMug: Pose-Agnostic Robotic Functional Grasping via Observation–Action Canonicalization
@misc{qiu2026anymug,
title = {Pose-Agnostic Robotic Functional Grasping via Observation-Action Canonicalization},
author = {Le Qiu and Cole Harrison and Jiankai Sun and Yao Liu and Suning Huang and Qianzhong Chen and Yang You and Marco Pavone},
year = {2026},
eprint = {2606.21148},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2606.21148}
}
A single closed-loop policy grasps mugs by the handle across shapes and poses by canonicalizing observations and actions in a mug-centric frame. Trained only in simulation, it transfers zero-shot to a real Franka Panda.

@misc{zhan2026sentinel,
title = {SENTINEL: A Multi-Level Formal Framework for Safety Evaluation of Foundation Model-based Embodied Agents},
author = {Simon Sinong Zhan and Philip Wang and Yao Liu and Yiyan Peng and Zinan Wang and Qineng Wang and Zhian Ruan and Xiangyu Shi and Xinyu Cao and Frank Yang and Zhenyang Ni and Kangrui Wang and Ruohan Zhang and Huajie Shao and Manling Li and Qi Zhu},
year = {2026},
eprint = {2510.12985},
archivePrefix = {arXiv},
primaryClass = {cs.AI},
url = {https://arxiv.org/abs/2510.12985}
}
SENTINEL specifies safety requirements for embodied LLM agents in temporal logic and checks them at the semantic, plan, and trajectory levels. Verification in VirtualHome and AI2-THOR flags unsafe states, ordering violations, and timed hazards with interpretable counterexamples.

@misc{liu2026dispatch,
title = {DISPATCH -- Decentralized Informed Spatial Planning and Assignment of Tasks for Cooperative Heterogeneous Agents},
author = {Yao Liu and Sampad Mohanty and Elizabeth Ondula and Bhaskar Krishnamachari},
year = {2026},
eprint = {2511.17915},
archivePrefix = {arXiv},
primaryClass = {cs.MA},
url = {https://arxiv.org/abs/2511.17915}
}
Decentralized policies for fair and efficient task assignment among heterogeneous agents under partial observability, learned toward Eisenberg–Gale equilibrium allocations with attention-based GNN message passing.

Quiet Multi-Agent Systems via Predictive World Modeling and Selective Communications
A conditional diffusion model infers global state from local views, and its uncertainty decides when robots need to communicate. The predictions guide MARL policies on NASA ISRU-style exploration and transport tasks.

An Efficient Distributed Inference Framework for IoT Systems with Byzantine Fault Detection
@inproceedings{zhou2025distributed,
title = {An Efficient Distributed Machine Learning Inference Framework with Byzantine Fault Detection},
author = {Zhou, Xuan and Mohan, Utkarsh and Liu, Yao and Beerel, Peter},
booktitle = {Proceedings of the Great Lakes Symposium on VLSI 2025},
series = {GLSVLSI '25},
pages = {56--63},
publisher = {ACM},
year = {2025},
doi = {10.1145/3716368.3735230}
}
A low-cost defense for LLM inference split across edge devices. Each device also computes a small, randomly chosen set of redundant activations, and the next layer flags any mismatch. We derive how many inferences detection takes and validate it on GPT-2 Small: with 10% overhead and 4 of 8 devices in a layer compromised, all malicious devices are identified in 39.42 inferences on average.

Let's Roll: Synthetic Dataset Analysis for Pedestrian Detection Across Different Shutter Types
@inproceedings{hu2024letsroll,
title = {Let's Roll: A Synthetic and Real Dataset for Pedestrian Detection Across Different Shutter Types},
author = {Hu, Yue and Datta, Gourav and Beerel, Kira and Liu, Yao and Beerel, Peter},
booktitle = {2024 IEEE Workshop on Signal Processing Systems (SiPS)},
pages = {113--118},
publisher = {IEEE},
year = {2024},
doi = {10.1109/SiPS62058.2024.00028}
}
Can detectors skip rolling-shutter correction? We built a dataset of 21,258 frames, in paired global- and rolling-shutter images, from Unreal Engine 5 and real scenes in downtown Los Angeles. YOLOv8 fine-tuned on rolling-shutter images nearly matches global-shutter performance at IoU 0.5, and adding synthetic data to the small real set improves recall by 53.1% and mAP@0.5 by 30.3% on real rolling-shutter data.
Predictive Modeling in AUV Navigation: A Perspective from Kalman Filtering
@misc{tang2026auv,
title = {Predictive Modeling in AUV Navigation: A Perspective from Kalman Filtering},
author = {Zizhan Tang and Yao Liu and Jessica Liu},
year = {2026},
eprint = {2603.27422},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2603.27422}
}
Multi-buoy TDoA measurements are fused with an adaptively tuned Kalman filter to track a submersible through communication loss. Filtered tracks yield a search heatmap, with meter-level accuracy in simulation.

Adaptive Online Pricing for Arrow–Debreu Markets
Online algorithms that compute Arrow–Debreu equilibrium prices as agents arrive with incomplete information, with O(log n) regret and O(1) constraint violation.
Selected Honors
- UC Berkeley EECS Ph.D. Fellowship, 2026
- Gold Medal, Kaggle LMSYS Chatbot Arena Human Preference Predictions, 2024
- Ming Hsieh Institute Undergraduate Scholar, USC (top 5 research scholars in EE), 2024
- Discover Scholar Distinction, USC, 2025
- Honorable Mention, Mathematical Contest in Modeling (COMAP), 2024
- Provost Research Scholarship, USC, 2024
- W.V.T. Rusch Engineering Honors Degree, USC, 2023
- Richardson CURVE Award, USC, 2022–2024
- Academic Achievement Award, USC, 2024–2025
- Dean's List, USC, 2022–2025
Service & Teaching
- Conference Reviewer: CoRL 2026, IEEE IV 2026
- Course Producer: EE 202 Linear Circuits, USC, Fall 2023 – Spring 2025
- Grader: EE 301 Linear Systems and MATH 245, USC, Fall 2024