Zihang Wang

Robot manipulation policies don't require architectural overhauls to survive visual distribution shifts—explicitly injecting language-modulated local feature maps into standard diffusion transformers doubles real-world bimanual task success.

MeshPriorDiT achieves a groundbreaking 75% reduction in prediction error for cloth dynamics by integrating local and global modeling strategies.

Achieving a dramatic reduction in joint-angle prediction error, the H-PAC hand demonstrates that underactuated systems can achieve high precision without complex retuning for different tasks.

Controller performance in edge clusters can swing dramatically based on workload intensity, with a deep reinforcement-learning approach losing to a simple heuristic by 29 percentage points under heavy load.

AIF-Router achieves stable online learning for adaptive AI service orchestration, even in unpredictable edge environments, showcasing the potential of Active Inference in real-world applications.