My research focuses on scalable robot learning. I study how to collect and learn from diverse demonstrations, use simulation as a scalable source of experience for real-world robots, and organize long demonstrations into training data for robot planning.
Under review, 2026
Built on AXIS, this benchmark studies how to use demonstration quality in VLA pretraining through controlled experiments on replayable data with simulator-grounded quality labels. Episode-level conditioning improves mean success across all three tested models. We also provide a real-world BC baseline pretrained only on workspace-aligned simulation data.
AXIS lets communities collect simulated robot demonstrations in a browser using everyday devices. Camera views, robot actions, and simulator states support replay and re-rendering for robot learning.
NeurIPS 2025
RDD turns long demonstrations into planner training segments matched to a robot’s learned skills. It retrieves similar examples from the robot policy’s training data to choose the segment boundaries.
| Oct 2026 | Release AXIS-Bench code and datasets are available. |
|---|---|
| Oct 2026 | Paper The AXIS-Bench paper is online. |
| Jul 2026 | Paper The AXIS preprint and project page are online. |
| Oct 2025 | Release RDD code is available, with the demonstration decomposition algorithm and example scripts. |
|---|---|
| Sep 2025 | Paper RDD is accepted to NeurIPS 2025. |
| Jun 2025 | Paper UniOcc is accepted to ICCV 2025. |
| Apr 2025 | Release UniOcc evaluation code is available for occupancy prediction and forecasting. |
| Jul 2024 | Experience Joined TASL as a Ph.D. student. |
ICCV 2025
A shared data format and evaluation framework for occupancy prediction and forecasting across real and simulated driving datasets.
arXiv preprint, 2025
Prioritizes object regions during point-cloud compression to preserve the geometry needed for 3D detection.
IEEE Transactions on Mobile Computing, 2024
Task-aware video compression preserves segmentation-relevant detail while adapting to available bandwidth.
IEEE INFOCOM 2023
Meta-learning enables streaming policies to adapt rapidly to changing network conditions.
Ph.D. in Machine Learning, Georgia Institute of Technology
Advisor: Prof. Jiachen Li
Ph.D. studies in Computer Science, University of California, Riverside
Transferred to Georgia Tech
M.Eng. in Information and Communication Engineering, Huazhong University of Science and Technology
Research Lead, 
Axis Robotics collaboration
Reviewer: ICRA, IEEE T-RO, ICLR, NeurIPS, CoRL.
