Paper-Conference | Zunzhe Zhang

Generative Control as Optimization: Time Unconditional Flow Matching for Adaptive and Robust Robotic Control

Sun, 26 Apr 2026 00:00:00 +0000

GeCO addresses a structural inefficiency in diffusion and flow-matching policies: a fixed inference schedule spends the same computation on easy and difficult states. The method treats action synthesis as optimization over a learned stationary velocity field, allowing adaptive computation at test time.

In addition to improving the success-latency trade-off, the field norm after optimization serves as an intrinsic OOD indicator. This makes the approach useful not only as a replacement for flow-matching action heads in VLA policies, but also as a mechanism for safer robotic deployment.

Zunzhe Zhang and Runhan Huang contributed equally to this work.

Virtual Community: An Open World for Humans, Robots, and Society

Sun, 08 Feb 2026 00:00:00 +0000

Virtual Community studies what happens when humans and robots coexist in shared, physically grounded communities. The platform combines scalable scene generation, real-world geospatial structure, grounded agent communities, and robot simulation powered by a physics engine.

The work introduces two benchmark settings: a Community Planning Challenge for open-world multi-agent reasoning, and a Community Robot Challenge for coordination among heterogeneous robots. Together, they make social and physical interaction testable at a scale that is difficult to study in isolated robotics benchmarks.

Ella: Embodied Social Agents with Lifelong Memory

Mon, 30 Jun 2025 00:00:00 +0000

Ella focuses on long-horizon embodied social behavior: agents must observe, remember, plan, and interact across days rather than solve isolated tasks. Its memory system organizes acquired knowledge in semantic form while preserving episodic traces of multimodal experience.

The system is evaluated in a dynamic 3D open world where multiple agents participate in social activities and are later tested on controlled tasks such as influence, leadership, and cooperation. The results show how structured memory can help foundation-model agents learn from observation and interaction.