Projects | Xuhong Wang

Selected Open Source Projects

A curated set of open source projects I actively drive, spanning trustworthy agent training, security tooling, data governance, multimodal content identification, and agent engineering plugins.

Safactory

A scalable agent infrastructure for trustworthy autonomous intelligence, unifying evaluation, trajectory collection, and reinforcement learning training across OS, Android, Minecraft, embodied AI, and scientific discovery environments.

DataElf

A data security agent for offline-isolated deployment, supporting least-privilege output, local auditing, and risk detection to keep sensitive data from leaking across the workflow.

Mokai Security Toolbox

An open-source toolbox for full-lifecycle agent security, bringing together 150+ tools for risk simulation, trusted data, and evolutionary defense.

UniMark

A multimodal AIGC watermark and identification toolkit for text, image, audio, and video with visible marking and invisible watermarking.

ClawReverse

A native version-control plugin for OpenClaw with checkpoints, rollback, and branching to cut rework and reduce repeated token usage.

Other Exploratory Open Source Projects

Additional exploratory projects spanning embodied navigation, long-term visual memory retrieval, multimodal perceptual robustness, and curated resources for trustworthy embodied AI.

Awesome Trustworthy Embodied AI

A curated resource hub for safe, reliable, interpretable, and aligned embodied AI, organizing papers, benchmarks, datasets, and open source projects.

IDMR

An instance-driven multimodal retrieval project for long-term embodied memory, enabling precise visual correspondence and memory recall from instance queries plus text conditions.

NaviMaster

A unified project page for GUI interaction and real-world navigation, presenting the task setup, modeling framework, and multi-scenario demos in one place.

AOT

An adversarial reinforcement learning project for multimodal perceptual robustness, where an attacker-defender co-evolution loop generates visual adversarial curricula to improve robustness and reduce hallucinations.