In this work, we advocate using Bayesian techniques for inversely identifying material parameters for multiscale crystal plasticity models. Multiscale approaches for modeling polycrystalline materials may significantly reduce the effort necessary for characterizing such material models experimentally, in particular when a large number of cycles is considered, as typical for fatigue applications. Even when appropriate microstructures and microscopic material models are identified, calibrating the individual parameters of the model to some experimental data is necessary for industrial use, and the task is formidable as even a single simulation run is time consuming (although less expensive than a corresponding experiment). For solving this problem, we investigate Gaussian process based Bayesian optimization, which iteratively builds up and improves a surrogate model of the objective function, at the same time accounting for uncertainties encountered during the optimization process. We describe the approach in detail, calibrating the material parameters of a high-strength steel as an application. We demonstrate that the proposed method improves upon comparable approaches based on an evolutionary algorithm and performing derivative-free methods.
Abstract-Whole body controllers based on quadratic programming allow humanoid robots to achieve complex motions. However, they rely on the assumption that the model perfectly captures the dynamics of the robot and its environment, whereas even the most accurate models are never perfect. In this paper, we introduce a trial-and-error learning algorithm that allows whole-body controllers to operate in spite of inaccurate models, without needing to update these models. The main idea is to encourage the controller to perform the task differently after each trial by introducing repulsors in the quadratic program cost function. We demonstrate our algorithm on (1) a simple 2D case and (2) a simulated iCub robot for which the model used by the controller and the one used in simulation do not match.
In recent years there has been a growing interest in the field of dynamic walking and bio-inspired robots. However, while walking and running on a flat surface have been studied extensively, walking dynamically over terrains with varying slope remains a challenge. Previously we developed an open loop controller based on a central pattern generator (CPG). The controller applied predefined torque patterns to a compass-gait biped, and achieved stable gaits over a limited range of slopes. In this work, this range is greatly extended by applying a once per cycle feedback to the CPG controller. The terrain's slope is measured and used to modify both the CPG frequency and the torque amplitude once per step. A multi-objective optimization algorithm was used to tune the controller parameters for a simulated CB model. The resulting controller successfully traverses terrains with slopes ranging from +7° to -8°, comparable to most slopes found in human constructed environments. Gait stability was verified by computing the linearized Poincaré Map both numerically and analytically.
Recent advances in control of humanoid robots have resulted in bipedal gaits that are dynamically stable on moderately rough terrain but are still limited to a small range of slopes. Humanoid robots, like humans, can take advantage of quadruped gaits to greatly extend this range. Cleverly designed gaits can provide robustness to rough terrain without requiring extensive feedback. In this paper, we present a robust crab-walking framework that includes forward and backward crawling patterns, rotation patterns, and sit-down and recovery sequences. The latter are activated autonomously once the robot detects that it tipped over. The performance and robustness of each locomotion pattern are investigated over a wide range of slopes. Crab-walking is shown to be especially adept at crawling forward on steep downward slopes (up to -54°) and crawling backward on steep upward slopes (up to 18°). Finally, we demonstrate the framework's autonomous capabilities by crossing the rough terrain in DARPA's virtual robotics challenge.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.