K
About

Kevin Ziyang Xie

I am interested in building learning algorithms that can self-improve from the world’s feedback without human involvement.

I’m working on the frontier of artificial intelligence and robot learning, supervised by Chris Xiaoxuan Lu at UCL RoMA Lab, as well as Yasemin Bekiroglu for my dissertation.

I am currently enrolled in the MEng Robotics and AI program at UCL. I was an applied scientist intern at Microsoft. Before that, I worked for Linear Capital on outreach and industry analysis.

Projects
Decent-VLA: An Empirical Study and Open Testbed for Federated Fine-Tuning of Vision-Language-Action Models submitted to ICRA 2027
Zhekai Duan*, Kevin Ziyang Xie, Xinyu Tan, Shikai Geng, Chengxu Zhou, Ramana Kompella, Gaowen Liu, Chris Xiaoxuan Lu
federated-learning VLA robot-learning

We study federated fine-tuning of three pretrained VLA policies across the 40 simulated tasks of LIBERO and two real-robot deployments of two and three sites, covering the gap to centralized fine-tuning, the choice of federated parameter scope, the aggregation algorithm, robustness under evaluation-time distribution shift, and the cross-site heterogeneity that appears only on physical robots; we release Decent-VLA, the testbed behind the study.

  • Decent-VLA released on arXiv
  • Decent-VLA merged into NVIDIA FLARE
arXiv Video NVFlare
Sponsored by Cisco NVIDIA
CRL-Gen: Can unsupervised RL act as a universal demonstrator?
Kevin Ziyang Xie, Zhekai Duan
June 2026
contrastive-rl reinforcement-learning multi-arm-manipulation

All systems today (2026) learn from human demos, and we have little control in terms of how much data we can produce or know what kind of data to produce. This was my earlier attempt in building systems that can learn by interacting with the world, with the hope that compute can directly be transferred into capabilities without the intermediate step of manual data collection.

Blog