Yiming Gan bio photo




Google Scholar

I’m leading a research group in the Institute of Computing Technology, Chinese Academy of Sciences. We work on autonomous machines, an exciting area that may reform the way we live in the future. We hope to build autonomous machines that are more intelligent, more efficient and more reliable. Our work spans designing new algorithms, designing new computer systems, and hopefully, designing real robots.

We currently have following research directions and projects:

Efficient Embodied AI Systems

Robots will have the ability of thinking. The goal of this direction is to enable robots to accomplish long and complex tasks and finish it in real time. The fundamental blocks here are in two folds. First, although large laugnage models have shown ability to accomplish long horizon tasks, the gap between the output of language models and real physical world are huge. Second, the envolve of huge LLMs makes it impossible for the robots to finish its decision making chain in real time. We currently have several related projects.

  • Large brain and small brain decoupling algorithm design.
  • Reinforcement learning based skill learning.
  • Cloud and edge collaborative processing for real time robotic applications.
  • Workload characterization for robots.
  • Efficient inference chips for language models.

Humanoid Robot Control

Robots will have the ability of moving like humans. The goal of this direction is to build algorithms and dedicate accelerators for controlling humanoid robots. The problem lies in how to handle high degree of freedoms within dynamic environments, how to accelerate control algorithms and achieve real time. We currently have several projects.

  • Robust control algorithms.
  • Dedicated control accelerators.

Multi-agents Communication

Robots will have the ability of communicating with each others. Ultimately we will enter an era where agents could communicate with each others. They will also share information with infrastructures. We currently have several project on this dimension.

  • Energy efficient federate learning.
  • Selective encryption with compilation optimization.

If your interests align with our group and want to work with us, please read this