Guanya Shi 石冠亚

I am a fifth-year Ph.D. candidate in the Computing and Mathematical Sciences (CMS) department at Caltech, advised by Prof. Soon-Jo Chung and Yisong Yue. I also collaborate closely with Prof. Anima Anandkumar, Adam Wierman, and Kamyar Azizzadenesheli. I am broadly interested in the intersection of machine learning and control theory, spanning the entire spectrum from theory to real-world agile robotics.

I received a B.E. from Tsinghua University in 2017. I was also an ML research intern at NVIDIA AI algorithm group in 2020. I was awarded the Simoudis Discovery Prize at Caltech in 2020 and was named a Rising Star in Data Science by the University of Chicago in 2021.

CV  /  Email  /  Blog  /  Google Scholar  /  Twitter  /  Research Statement


I am currently on the 2021-2022 academic job market, looking for faculty openings in Robotics/Machine Learning/Control (CS/EE/ME/AA/DS). Please feel free to contact me if you are interested. I would also be excited to talk about my research in related seminars.

News: I am co-organizing Control Meets Learning, a virtual seminar series on the intersection of control and learning.

News: I opened a blog focusing on control theory, machine learning, and robotics. Check it out!

News: Two theory papers about meta-adaptive control and predictive control were accepted by NeurIPS 2021.

News: Our Neural-Swarm2 paper was accepted by IEEE Transactions on Robotics (see the close-proximity flight of 16 drones and Yahoo! news).

News: Our FastUQ paper (my intern project at NVIDIA) was accepted by ICRA 2021.

News: Two theory papers about competitive control and the regret analysis of MPC were accepted by NeurIPS 2020.

News: Our work on Neural Lander was accepted by ICRA 2019 and highlighted by Caltech front page.

Research and Selected Publications

My research interests are in the intersection of machine learning and control theory, spanning the entire spectrum from theory and foundations, algorithm design, to real-world applications in robotics and autonomy. To that end, my research has three goals: (1) bridge learning and control theory in a unified framework; (2) design reliable learning and control algorithms with formal guarantees; and (3) push the boundaries of agile robotic control with new capabilities. For the full and up-to-date publication list, please visit my Google Scholar page. (* indicates equal contributions.)
Neural-Control Family: Stable and Robust Deep-learning-based Nonlinear Control in Dynamic Environments [blog post]

Neural-Lander: Stable Drone Landing Control using Learned Dynamics
Guanya Shi, Xichen Shi, Michael O'Connell, Rose Yu, Kamyar Azizzadenesheli, Animashree Anandkumar, Yisong Yue, Soon-Jo Chung
International Conference on Robotics and Automation (ICRA), 2019
[arXiv] [video] [Caltech front page news] [Import AI highlight] [PyTorch highlight]

We present a novel deep-learning-based robust nonlinear controller for stable quadrotor control during landing. Our approach blends together a nominal dynamics model coupled with a DNN that learns the high-order interactions, such as the complex interactions between the ground and multi-rotor airflow. This is the first DNN-based nonlinear feedback controller with stability guarantees that can utilize arbitrarily large neural nets. Neural-Lander enables agile drone maneuvers very close to the ground.

Neural-Swarm: Heterogeneous Multi-Robot Control and Planning Using Learned Interactions
Guanya Shi, Wolfgang Hoenig, Xichen Shi, Yisong Yue, Soon-Jo Chung
International Conference on Robotics and Automation (ICRA), 2020
IEEE Transactions on Robotics (T-RO), 2021
[arXiv] [video] [Caltech news] [Yahoo! news]

Close-proximity control and planning are challenging due to the complex aerodynamic effects between multirotors. We proposed Neural-Swarm, a nonlinear decentralized stable learning-based controller and motion planner for close-proximity flight of heterogeneous multirotor swarms. We develop and employ heterogeneous deep sets to encode multi-vehicle interactions in an index-free manner, enabling better generalization. Neural-Swarm enables close-proximity flight with 24 cm minimum vertical distance for a heterogeneous aerial team with 16 robots.

Neural-Fly: Rapid Learning for Agile Flight in Strong Winds
Michael O'Connell*, Guanya Shi*, Xichen Shi, Kamyar Azizzadenesheli, Animashree Anandkumar, Yisong Yue, Soon-Jo Chung
Science Robotics, under review
[perliminary version on arXiv]

Deep learning has representation power but is often too slow to update onboard. On the other hand, adaptive control with linear parametric uncertainty can update as fast as the feedback control loop. We propose an online stable and robust adaptation method that treats outputs from a DNN as a set of basis functions capable of representing different environments. A novel Domain Adversarially Invariant Meta-Learning (DAIML) algorithm is developed to train the network. Neural-Fly enables agile flights in unknown time-variant wind conditions.

Foundations of Online Learning and Control Theory

Meta-Adaptive Nonlinear Control: Theory and Algorithms
Guanya Shi, Kamyar Azizzadenesheli, Michael O'Connell, Soon-Jo Chung, Yisong Yue
Neural Information Processing Systems (NeurIPS), 2021
[arXiv] [code & video]

We present an online multi-task learning approach for adaptive nonlinear control, Online Meta-Adaptive Control (OMAC). The goal is to control a nonlinear system subject to adversarial disturbance and unknown environment-dependent nonlinear dynamics, under the assumption that the environment-dependent dynamics can be well captured with some shared representation. OMAC provides the first non-asymptotic end-to-end convergence guarantee for multi-task control. OMAC can also be integrated with deep representation learning.

Online Optimization with Memory and Competitive Control
Guanya Shi*, Yiheng Lin*, Soon-Jo Chung, Yisong Yue, Adam Wierman
Neural Information Processing Systems (NeurIPS), 2020
[arXiv] [NeurIPS video]

We presents competitive algorithms for a novel class of online optimization problems with memory. We consider a setting where the learner seeks to minimize the sum of a hitting cost and a switching cost that depends on the previous p decisions. The proposed approach, Optimistic ROBD, achieves aconstant, dimension-free competitive ratio. Further, we show a connection between online optimization with memory and online control with adversarial disturbances. This connection leads to a new constant-competitive policy for a rich class of online control problems.

The Power of Predictions in Online Control
Chenkai Yu, Guanya Shi, Soon-Jo Chung, Yisong Yue, Adam Wierman
Neural Information Processing Systems (NeurIPS), 2020
[arXiv] [NeurIPS video]

We study the impact of predictions in online LQR control with both stochastic and adversarial disturbances in the dynamics. In both settings, we characterize the optimal policy and derive tight bounds on the minimum cost and dynamic regret. Our analysis shows that the conventional MPC approach is a near-optimal policy in both settings. Specifically, for length-T problems, MPC requires O(logT) predictions to reach O(1) dynamic regret, which matches our lower bound on the required prediction horizon for constant regret. This result gives the first non-asymptotic guarantee for MPC.

Perturbation-Based Regret Analysis of Predictive Control in LTV Systems
Yiheng Lin*, Yang Hu*, Haoyuan Sun*, Guanya Shi*, Guannan Qu*, Adam Wierman
(Spotlight, <3% of Submissions) Neural Information Processing Systems (NeurIPS), 2021
[arXiv]

We study predictive control in linear time-varying (LTV) systems, and the costs are also time-varying. At each time step, the controller receives the exact predictions of costs, dynamics, and disturbances for the future k time steps. We show that MPC achieves a dynamic regret of O(λkT), where λ<1 is a positive constant. We also show that MPC obtains the first competitive bound for the control of LTV systems: 1+O(λk). Our results are derived using a novel proof framework based on a perturbation bound that characterizes how a small change to the system parameters impacts the optimal trajectory.

Uncertainty Quantification and Safe Explorations in Dynamical Systems

Fast Uncertainty Quantification for Deep Object Pose Estimation
Guanya Shi, Yifeng Zhu, Jonathan Tremblay, Stan Birchfield, Fabio Ramos, Animashree Anandkumar, Yuke Zhu
International Conference on Robotics and Automation (ICRA), 2021
[arXiv] [project website] [code] [Nvidia developer blog]

Deep learning-based object pose estimators are often unreliable and overconfident especially with sim2real transfer. In this work, we propose a simple, efficient, and plug-and-play UQ method for 6-DoF object pose estimation. We ensemble pre-trained models with different neural network architectures and/or training data sources, and compute their average pairwise disagreement against one another to obtain the uncertainty quantification. Our UQ method yields much stronger correlations with pose estimation errors than the baselines in three tasks. Moreover, in a real robot grasping task, our method increases the grasping success rate from 35% to 90%.

Robust Regression for Safe Exploration in Control
Anqi Liu, Guanya Shi, Soon-Jo Chung, Animashree Anandkumar, Yisong Yue
Conference on Learning for Dynamics and Control (L4DC), 2020
[arXiv]

We study the problem of safe learning and exploration in sequential control problems. The goal is to safely collect data samples from operating in an environment, in order to learn to achieve a challenging control goal (e.g., an agile maneuver close to a boundary). We present a deep robust regression model that is trained to directly predict the uncertainty bounds for safe exploration. We derive generalization bounds for learning under domain shift and connect them with safety and stability bounds in control. Our approach outperforms the conventional Gaussian process (GP) based safe exploration in settings where it is difficult to specify a good GP prior.

Chance-Constrained Trajectory Optimization for Safe Exploration and Learning of Nonlinear Systems
Yashwanth Kumar Nakka, Anqi Liu, Guanya Shi, Animashree Anandkumar, Yisong Yue, Soon-Jo Chung
IEEE Robotics and Automation Letters (RA-L), 2020
[arXiv] [blog]

We present a new approach for optimal motion planning with safe exploration that integrates chance-constrained stochastic optimal control with dynamics learning and feedback control. We derive an iterative convex optimization algorithm that solves an Information-cost Stochastic Nonlinear Optimal Control problem (Info-SNOC). The optimization objective encodes both optimal performance and exploration for learning, and the safety is incorporated as distributionally robust chance constraints. We prove the safety of rollout from our exploration method and reduction in uncertainty over epochs, thereby guaranteeing the consistency of our learning method. Our approach has higher success rate in ensuring safety when compared to a deterministic trajectory optimization approach.


Misc.

I love playing basketball, soccer and MOBA games. I am also very interested in photography, hiking, travelling and cooking. Here are some photos taken by me. Feel free to email me if you are interested in these photos.

prl prl
A drone flying in the Caltech Real Weather Wind Tunnel Caltech Beckman Auditorium
prl prl
Beijing National Stadium Catalina Island
prl prl
Wudaokou, Beijing Tokugawaen, Nagoya, Japan
prl prl
Winter Tsinghua Yosemite, California


Based on this website.