Junliang Zhang

Junliang Zhang


Research Focus & Metrics

Academic Career Length

1Year (2026-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
16 November 2026 in Preprints.org
Verified
Preprint
Open access
PDF file
MSMeilin Suo
Meilin Suo
JXJun Xie
Jun Xie
JZJunliang Zhang
Junliang Zhang
HLHuanxu Lin
Huanxu Lin
XMXiaoguang Ma
Xiaoguang Ma
XDXiangyu Dong
Xiangyu Dong
Recent advances in foundation models have enabled robotic agents to reason, plan, and adapt their actions through physical interaction. Yet successful execution does not necessarily translate into lasting improvement across tasks. Recursive self-improvement (RSI) offers a promising paradigm for robotic agents to continuously refine their capabilities through experience, but requires preserving functional knowledge, adapting execution to changing physical conditions, and retaining reusable corrections. We introduce LoopROS, a dual-loop robotic agent framework that couples functionally constrained execution with experience-driven improvement. Within each episode, functional constraint-guided in-context learning identifies essential task relations from demonstrations and context, while visual feedback-guided segmentwise trajectory tuning adapts reference motions through observation-grounded corrections and verification. Across episodes, execution feedback informs revisions to persistent skills, parameters, and executable tools, enabling subsequent executions to reuse corrections under relevant conditions. The LoopROS achieves an 82.0% success rate and a mean score of 88.90, ranking No. 1 in RoboDojo, among the compared methods. We further complete 30 successful trials in three real robot manipulation tasks, elucidating its great potential as a backbone robotic RSI.