Learning To Adapt For Intelligent Robot Behavior PDF Download

The field of robotics has been rapidly evolving in recent years, and robots are being used in an ever-increasing number of applications, from manufacturing to healthcare to household chores. One of the key challenges in robotics is enabling robots to perform complex manipulation tasks in unstructured and dynamic environments. While there have been significant advances in robot learning and control, many existing approaches are limited by their reliance on pre-defined motion primitives or generic models that do not account for the specific characteristics of individual users, other cooperative agents or the interacting objects. In order to be effective in these various settings, robots need to be able to adapt to different tasks and environments, and to interact with different types of agents, such as humans and other robots. This thesis investigates learning approaches for enabling robots to adapt their behavior in order to achieve intelligent robot behavior. In the first part of this thesis, we focus on enabling robots to better adapt to humans. We start by exploring how to leverage different sources of data to achieve personalization for human users. Firstly, we investigate how humans prefer to teleoperate assistive robot arms using low-dimensional controllers, such as joysticks. We present an algorithm that can efficiently develop personalized control for assistive robots. Here the data is obtained by initially demonstrating the behavior of the robot and then query the user to collect their corresponding preferred teleoperation control input from the joysticks. Subsequently, we delve into the exploration of leveraging weaker signals to infer information from agents, such as physical corrections. Experiment results indicate that human corrections are correlated and reasoning over these corrections together achieves improved accuracy. Finally, instead of only adapting to a single human user, we investigate how robots can more efficiently cooperate with and influence human teams by reasoning and exploiting the team structure. We apply our framework to two types of group dynamics, leading-following and predator-prey, and demonstrate that robots can first develop a group representation and utilize this representation to successfully influence a group to achieve various goals. In the second part of this thesis, we extend our investigation from human users to robot agents. We tackle the problem of how decentralized robot teams can adapt to each other by observing only the actions of other agents. We identify the problem of an infinite reasoning loop within the team and propose a solution by assigning different roles, such as "speaker" and "listener, " to the robot agents. This approach enables us to treat observed actions as a communication channel, thereby achieving effective collaboration within the decentralized team. Moving on to the third part of this thesis, we explore the topic of adapting to different tasks by developing customized tools. We emphasize the critical role of tools in determining how a robot interacts with objects, making them important in customizing robots for specific tasks. To address this, we present an end-to-end framework to automatically learn tool morphology for contact-rich manipulation tasks by leveraging differentiable physics simulators. Finally, we conclude the thesis by summarizing our efforts and discussing future directions.