Papers

Generative Control as Optimization (GeCO) reframes flow-matching action generation for robotic control as adaptive iterative optimization. Instead of following a fixed integration schedule, GeCO learns a stationary velocity field in action-sequence space so inference can stop early for simple states and refine longer for difficult states. The same stationary geometry also provides a training-free signal for detecting out-of-distribution states, supporting safer deployment of Vision-Language-Action policies.

@inproceedings{zhang2026generative,
  title={Generative Control as Optimization: Time Unconditional Flow Matching for Adaptive and Robust Robotic Control},
  author={Zhang, Zunzhe and Huang, Runhan and Liu, Yicheng and Zhu, Shaoting and Mou, Linzhan and Zhao, Hang},
  booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
  year={2026},
  url={https://arxiv.org/abs/2603.17834},
  doi={10.48550/arXiv.2603.17834}
}

Virtual Community is an open-world platform for studying embodied social intelligence across humans, robots, and society. Built on real-world grounded 3D scenes and a multi-agent physics simulator, it supports social interaction, robot-human coexistence, and large-scale community generation. The paper introduces benchmark challenges for community planning and heterogeneous robot collaboration, highlighting open problems in high-level planning and low-level cooperative control.

@inproceedings{zhou2026virtual,
  title={Virtual Community: An Open World for Humans, Robots, and Society},
  author={Zhou, Qinhong and Zhang, Hongxin and Lin, Xiangye and Zhang, Zheyuan and Chen, Yutian and Liu, Wenjun and Zhang, Zunzhe and Chen, Sunli and Fang, Lixing and Lyu, Qiushi and Sun, Xinyu and Yang, Jincheng and Wang, Zeyuan and Dang, Bao Chi and Chen, Zhehuan and Ladia, Daksha and Dang, Quang Vinh and Liu, Jiageng and Gan, Chuang},
  booktitle={International Conference on Learning Representations},
  year={2026},
  url={https://arxiv.org/abs/2508.14893},
  doi={10.48550/arXiv.2508.14893}
}

Ella is an embodied social agent designed for lifelong learning in a 3D open world. It combines foundation models with a structured, non-parametric long-term multimodal memory system, including name-centric semantic memory and spatiotemporal episodic memory. Through extended observation and social interaction, Ella retrieves relevant memories, plans daily behavior, builds relationships, and adapts in multi-agent community settings.

@inproceedings{zhang2025ella,
  title={Ella: Embodied Social Agents with Lifelong Memory},
  author={Zhang, Hongxin and Zhang, Zheyuan and Wang, Zeyuan and Zhang, Zunzhe and Fang, Lixing and Zhou, Qinhong and Gan, Chuang},
  booktitle={ICML MAS Workshop},
  year={2025},
  url={https://arxiv.org/abs/2506.24019},
  doi={10.48550/arXiv.2506.24019}
}