Guanya Shi
Carnegie Mellon University, PA·Engineering
Humanoid robots and other autonomous machines must choose good actions in real time, in messy and unpredictable settings. Their physics is often highly nonlinear, and moments of contact, such as a foot striking the ground or a hand grasping an object, can even make it discontinuous.
- Fields
Engineering
Computer Science
Mathematics & Statistics
- Topics
sampling-based optimal control
model predictive control
humanoid robots
- Grants
- 1 known grant, through Aug 2028
What they work on
- Develop finite-time and long-run convergence theories that show when sampling-based optimal control reaches the best solution and how its performance scales with algorithm and system parameters.
- Design more efficient sampling-based control algorithms using automated hyperparameter optimization and a diffusion-style annealing strategy that speeds up exploration while preserving convergence.
- Combine sampling-based controllers with learned value functions, nominal policies and dynamics models, creating a subroutine for model-based reinforcement learning.
- Validate the methods in real hardware experiments on humanoid whole-body locomotion and dexterous robotic manipulation.
More about the research
Classical optimal control, the mathematics of choosing the best actions over time, tends to be slow on such problems and can get stuck on poor solutions. Sampling-based optimal control (SBOC) offers a different approach. It simulates many random trial trajectories in parallel, often on GPUs, and keeps the best ones. SBOC works well in practice for robot planning and control, but it lacks theory: there are no guarantees it will find the best answer, and little guidance on how to tune it.
Why it matters
Robots that walk, handle objects or operate on their own in unstructured settings need decision-making that is fast and reliable. Giving sampling-based control mathematical guarantees and clear design rules could make humanoid robots and other autonomous systems more dependable. It could also make them easier to build for real-world tasks where failures are costly.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
Closest labs
Maegan Tucker
Georgia Institute of Technology, GA·Engineering
Extend contraction theory, a mathematical test of whether nearby motions of a system converge, to hybrid systems that mix smooth motion with sudden impacts. Use it to estimate regions of attraction for contact-rich robots.
Bruno Sinopoli
Arizona State University Campus Immersion, AZ·Engineering
Establish closed-loop stability guarantees for REAP solvers, optimization methods that return a safe, feasible answer even when stopped early, and develop ways to implement them on discrete-time digital hardware.
Mengdi Wang
Princeton University, NJ·Computer Science
Develop distributionally robust training methods for model-based reinforcement learning so that policies avoid unsafe states even when the real environment differs from the training data.
The school
Carnegie Mellon University: 28 labs on PI Pages, $505M a year on research, 346 research doctorates a year.
Reach out
A start for your email. Fill in the brackets.