I am a master's student at UCAS (2024–2027), advised by Prof. Xiangsheng Huang. My research focuses on VLA post-training and representation stability, cross-task generalization, and real-robot learning.
Research Interests
VLA post-training and representation stability
Cross-task generalization and real-robot learning
WAM and deployment-aware model adaptation
Publications
FiberTune: Preserving Action-Fiber Visual Residuals in Vision-Language-Action Fine-Tuning
Haihao Lin (first & corresponding author). Accepted at CoRL 2026.
[arXiv]
[PDF]
[Project]
FiberTune preserves transferable visual structure during VLA fine-tuning without adding inference-time overhead. It improves over task-loss-only fine-tuning across six matched CALVIN / LIBERO settings and on physical SO-101, where task success increased from 72.7% to 78.1%.
Research & Systems
Current experiments span VLA and WAM, using FastWAM and related models to study cross-task generalization and real-robot learning.
Designed and implemented autonomous docking for charging, cliff detection and avoidance, and a corresponding docking simulation system for a commercial cleaning robot, followed by ROS2 deployment and system-level closed-loop validation.
Open Source
I was an early key contributor to
MaaAssistantArknights
(20k+ GitHub stars), where I developed C++ core modules, WPF GUI features, and a task-planning algorithm for mutual-exclusion constraints.
I also contributed to HoshinoBot.