RI: CAREER: Task-Oriented Model Identification for Robust Robotic Manipulation

点击翻译按钮获取中文摘要

Robots typically rely on models of their mechanical components and of the objects in their surroundings to perform their tasks reliably. The models describe the shapes and the mechanical properties of objects. The models are used to simulate different actions that a robot can perform, and the actions with the best forecasted outcomes are selected for execution in the real environment. In practice, the forecasted outcomes are often different from the real outcomes due to the inaccuracies of the models, this difference is what is called the reality gap. Manually-designed models are inherently inaccurate. While this problem is less pronounced in industrial robots that typically operate in closed, structured and controlled environments, it severely limits the deployment of robots to open environments where they constantly encounter novel objects with unknown or uncertain models. For example, an assistant robot in a repair shop needs to manipulate various tools and operate on new objects everyday. The goal of this project is to develop automated and data-driven object modeling methods that will allow robots to build geometric and mechanical models of objects on the fly while manipulating them cautiously. Anticipated improvements have the potential for impact in several application areas, such as job shops that require high flexibility in product engineering, household robotics, and debris removal in rescue operations. This project fosters these potentials by creating a new course and textbook in robot learning, and releasing general purpose object modeling tools, while organizing museum exhibitions that will expose automated object modeling and manipulation techniques to a wider audience. Additionally, the project seeks to involve undergraduates in research activities at Rutgers, The State University of New Jersey, which serves a diverse student population.The approach pursued in this project is to automatically generate and gradually fine-tune mechanical models of objects by searching for models that minimize the gaps between simulation and reality. Specifically, the goal here is not to identify the most accurate model of an object, but rather to infer models that are sufficiently accurate to perform a given manipulation task. Therefore, the automated modeling process is strongly guided by the given manipulation task, unnecessary computational modeling efforts are thus avoided. The main technical objectives of this project are to: 1) Provide theoretical guarantees on the performance of control techniques using imperfect models inferred from data. 2) Develop black-box Bayesian optimization tools for inferring models of objects from limited vision and interaction data. 3) Develop white-box model identification tools using differentiable 3D renderers and physics engines. 4) Demonstrate the developed methods on a diverse range of tasks related to manipulating unknown objects in cluttered environments.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.