Gerardo Bledt/Donghyun Kim (Sangbae Kim) - Postdoctoral Associates, MIT [Chair: Marko Bjelonic; Moderator: Aaron Ames]
Title: Regularized Predictive Control: Meaningful, Simple Optimization for Robust, Dynamic Robots
Abstract: Designing controllers for complex dynamic robots in unstructured terrains is a challenging, open-ended task. Optimization-based predictive controllers have proven to be powerful in producing highly robust, dynamic behaviors in legged systems. However, since "optimal solutions" are only as good as the underlying dynamics model, constraints, and cost function, we must design meaningful optimizations that produce the desired behaviors, while remaining computationally tractable. Regularized Predictive Control (RPC) is a novel nonlinear optimization-based predictive controller that optimizes robot states, footstep locations, and ground reaction forces over a future prediction horizon in real-time. The control model is presented as a highly simplified, rough feasibility constraint, while simple heuristics are embedded directly into the cost function through regularization to improve performance. The challenge of designing meaningful heuristics is tackled by a combination of expert design and data-driven analytics for extracting and fitting simple models to simulation results. These simple models are able to simultaneously capture the major effects on the system dynamics, while reducing the complexity of the controller. The task of control design is ultimately to find a controller that robustly achieves the desired performance, rather than to describe the system accurately. Often times, a "good enough" solution is better than an analytically complex or optimal one.
Bio: Gerardo is currently a Postdoc at the Biomimetic Robotics Lab at MIT with Professor Sangbae Kim working on dynamic legged robots where he received his PhD in Robotics this year for his research titled, "Regularized Predictive Control Framework for Robust Dynamic Legged Locomotion." The work dealt with designing a real-time nonlinear optimization-based controller for legged robots. RPC allows direct embedding of known physics and extracted data-driven models into the optimization as simple regularization heuristics. The result was a quadruped robot capable of highly dynamic maneuvers that is robust to unpredictable terrain and large disturbances. Research interests include intelligent control algorithms for dynamic systems, data-driven methods for extracting simple models describing complex systems, planning motions under uncertainty, and many more aspects of robotics in general. Gerardo also holds Masters degrees in Mechanical Engineering, Electrical Engineering & Computer Science from MIT and a Bachelors Degree in Mechanical Engineering from Virginia Tech.
Title: Whole-Body Impulse Control and Vision Aided Dynamic Walking
Bio: Donghyun Kim is currently a Postdoctoral Associate at the Massachusetts Institute of Technology and will join in College of Information and Computer Science department at the University of Massachusetts Amherst as an Assistant Professor. Donghyun's primary research area is in the dynamic locomotion of legged systems with a focus on the development of control architectures and their experimental validation. During his Ph.D. at UT Austin, Donghyun developed controllers for passive-ankle biped robots. At MIT, he developed controllers for high speed running of quadruped robots and demonstrated the Mini-Cheetah robot running up to 3.7 m/s. He is now extending his research area to a perception-based high-level decision algorithm to push forward robots' athletic intelligence.
Abstract: Dynamic legged locomotion is a challenging topic because of the lack of established control architecture which can handle aerial phases, short stance times, and high-speed leg swings. We formulated a controller combining whole-body control (WBC) and model predictive control (MPC) to tackle those issues. Unlike existing WBCs, which attempt to track commanded body trajectories, our controller is focused more on the reaction force command, which allows it to accomplish high-speed dynamic locomotion with aerial phases, which is up to 3.7 m/s with a Mini-Cheetah robot. On top of the advanced locomotion controller, we added vision sensors to extend a Mini-Cheetah’s terrain coverage. Rough terrain locomotion of a small robot has unique difficulties such as limited space for sensors, limited obstacle clearance, and scaled-down walking speed, but we overcame the problems with efficient sensor integration and exploitation of dynamic walking and jumping.
ICRA 2020 Workshop website: https://sites.google.com/view/leggedrobotworkshop2020