Simin Liu

I'm a final year PhD student at CMU's Robotics Institute. I work on control and motion planning for robot manipulators and drones. In general, my goal is to create safe, useful robots. I'm grateful to be advised by Changliu Liu and John Dolan and supported by the Qualcomm Innovation Fellowship.

Before my PhD, I received my BS with high honors in EECS and Math at UC Berkeley. There, I did research with Sergey Levine on adaptive control for legged locomotion.

I will be doing a research internship at the Boston Dynamics AI Institute this fall. I am working on planning for manipulation with Jiuguang Wang and Tao Peng. If you are in Boston and want to meet up, please reach out!

Email me at siminliu at andrew dot cmu dot edu

CV / Google Scholar / GitHub / LinkedIn

profile photo

Research

project image

Synthesis and Verification of Robust-Adaptive Safe Controllers


Simin Liu*, Kai Yun*, John M Dolan, Changliu Liu
European Controls Conference, 2024, 2023
arxiv /

One major drawback of prior work is that it cannot satisfactorily handle model uncertainty. We designed an optimization algorithm for generating robust-adaptive safe controllers, which account for uncertainty without producing over-conservative behavior. Ultimately, our controller allows for 55% better control performance while still ensuring safety as compared to popular baselines.

project image

Safety Index Synthesis via Sum-of-Squares Programming


Weiye Zhao, Tairan He, Tianhao Wei, Simin Liu, and Changliu Liu
American Controls Conference, 2023
arxiv /

Safety indices, also called control barrier functions (CBFs), are popularly used for safe reactive control. However, for all but the simplest systems, it’s hard to find a valid CBF in the first place. We devise an optimization algorithm that can automatically generate a CBF for any moderate-sized, polynomial-equivalent system. This includes simplified vehicles and manipulators.

project image

Safe Control Under Input Limits with Neural Control Barrier Functions


Simin Liu, Changliu Liu, John Dolan
Conference on Robot Learning, 2022
arxiv / code / slides /

The work below and its alternatives can only scale to 5-7 D systems. We created a much more scalable technique using adversarial training of neural CBFs. Essentially, our method trades the theoretical guarantees of safety for scalability and strong empirical guarantees (>99% safe). Currently, this class of methods has been shown to scale to >20D. This includes complex systems, like balancing quadrotors and many-linked manipulators.

project image

Learning to Adapt in Dynamic, Real-World Environments Through Meta-Reinforcement Learning


Anusha Nagabandi*, Ignasi Clavera*, Simin Liu, Ronald Fearing, Pieter Abbeel, Sergey Levine, Chelsea Finn
International Conference on Learning Representations, 2018
website / arxiv / code /

The system dynamics of robots can sometimes change online (e.g. as the environment changes). To deal with this, we consider adapting a neural dynamics model online. We propose to meta-learn a model initialization and update it online with gradient descent. This scheme can adapt very well after just a few samples of data. We show that a legged millirobot using this scheme can adapt to dynamics that change with terrain, payload, and wear-and-tear.


Hobbies

I like outdoor activities, drawing, ceramics, reading fiction, swimming, running.



Design and source code from Leo Keselman