Auto-Tuning

In the past year, I have been working on three separate works that employ an Auto-Tuning approach to calibrate user-defined parameters. These works span diverse sets of applications: hybrid control and planning for a quadruped robot, force control for robotic wall climbing, and learning residual errors for real-to-sim applications. One of these works is published while the other two are currently under review (see title and short description of each work below), and will be shared soon!

Auto-Tuning of Controller and Online Trajectory Planner for Legged Robots

Summary: An auto-tuning method is demonstrated on feedback controllers and online trajectory planners to achieve robust locomotion of a legged robot. Cost function weights of a Model Predictive Controller and feedback gains of a swing controller are calibrated. Further, the auto-tuning approach is used to calibrate parameters of an online trajectory planner, where the height of a swing leg and robot's walking speed are optimized.

Authors: Alexander Schperberg, Stefano Di Cairano, and Marcel Menner

Project Information coming soon (Accepted to Robotics and Automation Letters Journal - June 2022)

Auto-Calibrating Admittance Controller for Robust Motion of Robotic Systems

Summary: We demonstrate a self-calibrating admittance control formulation that enables wrench control (force and torques) for both manipulation and locomotion tasks. Our controller is self-calibrating using an auto-tuning method with training objectives that facilitate controller robustness/adaptability during online operation (e.g., ensuring friction cone is satisfied, spring constants are updated online to increase robustness, and reference trajectories are tracked continuously).

Authors: Alexander Schperberg, Yuki Shirai, Xuan Lin, Yusuke Tanaka, and Dennis Hong

Project information coming soon (Under Review)

Real-to-Sim: Deep Learning with Auto-Tuning to Predict Residual Model Error

Residual model error is predicted between a simulator and the reality. The goal is to use real world data to improve the simulator model. Typically, this is a very challenging task because a substantial amount of training data is required -- this makes it difficult to train directly on hardware. However, we propose a neural network with auto-tuning approach, that can quickly converge on sparse amounts of data, a necessity for training on hardware directly.

Authors: Alexander Schperberg, Yusuke Tanaka, Feng Xu, Marcel Menner, and Dennis Hong