Organizational Unit:

Institute for Robotics and Intelligent Machines (IRIM)

Permanent Link

https://hdl.handle.net/1853/70657

Includes Organization(s)

Organizational Unit

Biorobotics and Human Modeling Lab

Organizational Unit

Humanoid Robotics Laboratory

Full item page

Publication Search Results

Now showing 1 - 10 of 73

A control theoretic perspective on learning in robotics

(Georgia Institute of Technology, 2015-12-16) O'Flaherty, Rowland Wilde

For robotic systems to continue to move towards ubiquity, robots need to be more autonomous. More autonomy dictates that robots need to be able to make better decisions. Control theory and machine learning are fields of robotics that focus on the decision making process. However, each of these fields implements decision making at different levels of abstraction and at different time scales. Control theory defines low-level decisions at high rates, while machine learning defines high-level decision at low rates. The objective of this research is to integrate tools from both machine leaning and control theory to solve higher dimensional, complex problems, and to optimize the decision making process. Throughout this research, multiple algorithms were created that use concepts from both control theory and machine learning, which provide new tools for robots to make better decisions. One algorithm enables a robot to learn how to optimally explore an unknown space, and autonomously decide when to explore for new information or exploit its current information. Another algorithm enables a robot to learn how to locomote with complex dynamics. These algorithms are evaluated both in simulation and on real robots. The results and analysis of these experiments are presented, which demonstrate the utility of the algorithms introduced in this work. Additionally, a new notion of “learnability” is introduced to define and determine when a given dynamical system has the ability to gain knowledge to optimize a given objective function.
Towards a terradynamics of legged locomotion on homogeneous and Heterogeneous granular media through robophysical approaches

(Georgia Institute of Technology, 2015-11-16) Qian, Feifei

The objective of this research is to discover principles of ambulatory locomotion on homogeneous and heterogeneous granular substrates and create models of animal and robot interaction within such environments. Since interaction with natural substrates is too complicated to model, we take a robophysics approach – we create a terrain generation system where properties of heterogeneous multi-component substrates can be systematically varied to emulate a wide range of natural terrain properties such as compaction, orientation, obstacle shape/size/distribution, and obstacle mobility within the substrate. A schematic of the proposed system is discussed in detail in the body of this dissertation. Control of such substrates will allow for the systematic exploration of parameters of substrate properties, particularly substrate stiffness and heterogeneities. With this terrain creation system, we systematically explore locomotor strategies of simplified laboratory robots when traversing over different terrain properties. A key feature of this proposed work is the ability to generate general interaction models of locomotor appendages with such complex substrates. These models will aid in the design and control of future robots with morphologies and control strategies that allow for effective navigation on a large diversity of terrains, expanding the scope of terramechanics from large tracked and treaded vehicles on homogeneous ground to arbitrarily shaped and actuated locomotors moving on complex heterogeneous terrestrial substrates.
Machine learning and dynamic programming algorithms for motion planning and control

(Georgia Institute of Technology, 2015-11-16) Arslan, Oktay

Robot motion planning is one of the central problems in robotics, and has received considerable amount of attention not only from roboticists but also from the control and artificial intelligence (AI) communities. Despite the different types of applications and physical properties of robotic systems, many high-level tasks of autonomous systems can be decomposed into subtasks which require point-to-point navigation while avoiding infeasible regions due to the obstacles in the workspace. This dissertation aims at developing a new class of sampling-based motion planning algorithms that are fast, efficient and asymptotically optimal by employing ideas from Machine Learning (ML) and Dynamic Programming (DP). First, we interpret the robot motion planning problem as a form of a machine learning problem since the underlying search space is not known a priori, and utilize random geometric graphs to compute consistent discretizations of the underlying continuous search space. Then, we integrate existing DP algorithms and ML algorithms to the framework of sampling-based algorithms for better exploitation and exploration, respectively. We introduce a novel sampling-based algorithm, called RRT#, that improves upon the well-known RRT* algorithm by leveraging value and policy iteration methods as new information is collected. The proposed algorithms yield provable guarantees on correctness, completeness and asymptotic optimality. We also develop an adaptive sampling strategy by considering exploration as a classification (or regression) problem, and use online machine learning algorithms to learn the relevant region of a query, i.e., the region that contains the optimal solution, without significant computational overhead. We then extend the application of sampling-based algorithms to a class of stochastic optimal control problems and problems with differential constraints. Specifically, we introduce the Path Integral - RRT algorithm, for solving optimal control of stochastic systems and the CL-RRT# algorithm that uses closed-loop prediction for trajectory generation for differential systems. One of the key benefits of CL-RRT# is that for many systems, given a low-level tracking controller, it is easier to handle differential constraints, so complex steering procedures are not needed, unlike most existing kinodynamic sampling-based algorithms. Implementation results of sampling-based planners for route planning of a full-scale autonomous helicopter under the Autonomous Aerial Cargo/Utility System Program (AACUS) program are provided.
Developing an engagement and social interaction model for a robotic educational agent

(Georgia Institute of Technology, 2015-11-16) Brown, LaVonda N.

Effective educational agents should accomplish four essential goals during a student's learning process: 1) monitor engagement, 2) re-engage when appropriate, 3) teach novel tasks, and 4) improve retention. In this dissertation, we focus on all of these objectives through use of a teaching device (computer, tablet, or virtual reality game) and a robotic educational agent. We begin by developing and validating an engagement model based on the interactions between the student and the teaching device. This model uses time, performance, and/or eye gaze to determine the student's level of engagement. We then create a framework for implementing verbal and nonverbal, or gestural, behaviors on a humanoid robot and evaluate its perception and effectiveness for social interaction. These verbal and nonverbal behaviors are applied throughout the learning scenario to re-engage the students when the engagement model deems it necessary. Finally, we describe and validate the entire educational system that uses the engagement model to activate the behavioral strategies embedded on the robot when learning a new task. We then follow-up this study to evaluate student retention when using this system. The outcome of this research is the development of an educational system that effectively monitors student engagement, applies behavioral strategies, teaches novel tasks, and improves student retention to achieve individualized learning.
Robots learning actions and goals from everyday people

(Georgia Institute of Technology, 2015-11-16) Akgun, Baris

Robots are destined to move beyond the caged factory floors towards domains where they will be interacting closely with humans. They will encounter highly varied environments, scenarios and user demands. As a result, programming robots after deployment will be an important requirement. To address this challenge, the field of Learning from Demonstration (LfD) emerged with the vision of programming robots through demonstrations of the desired behavior instead of explicit programming. The field of LfD within robotics has been around for more than 30 years and is still an actively researched field. However, very little research is done on the implications of having a non-robotics expert as a teacher. This thesis aims to bridge this gap by developing learning from demonstration algorithms and interaction paradigms that allow non-expert people to teach robots new skills. The first step of the thesis was to evaluate how non-expert teachers provide demonstrations to robots. Keyframe demonstrations are introduced to the field of LfD to help people teach skills to robots and compared with the traditional trajectory demonstrations. The utility of keyframes are validated by a series of experiments with more than 80 participants. Based on the experiments, a hybrid of trajectory and keyframe demonstrations are proposed to take advantage of both and a method was developed to learn from trajectories, keyframes and hybrid demonstrations in a unified way. A key insight from these user experiments was that teachers are goal oriented. They concentrated on achieving the goal of the demonstrated skills rather than providing good quality demonstrations. Based on this observation, this thesis introduces a method that can learn actions and goals from the same set of demonstrations. The action models are used to execute the skill and goal models to monitor this execution. A user study with eight participants and two skills showed that successful goal models can be learned from non- expert teacher data even if the resulting action models are not as successful. Following these results, this thesis further develops a self-improvement algorithm that uses the goal monitoring output to improve the action models, without further user input. This approach is validated with an expert user and two skills. Finally, this thesis builds an interactive LfD system that incorporates both goal learning and self-improvement and evaluates it with 12 naive users and three skills. The results suggests that teacher feedback during experiments increases skill execution and monitoring success. Moreover, non-expert data can be used as a seed to self-improvement to fix unsuccessful action models.
Developing robots that impact human-robot trust in emergency evacuations

(Georgia Institute of Technology, 2015-11-10) Robinette, Paul

High-risk, time-critical situations require trust for humans to interact with other agents even if they have never interacted with the agents before. In the near future, robots will perform tasks to help people in such situations, thus robots must understand why a person makes a trust decision in order to effectively aid the person. High casualty rates in several emergency evacuations motivate our use of this scenario as an example of a high-risk, time-critical situation. Emergency guidance robots can be stored inside of buildings then activated to search for victims and guide evacuees to safety. In this dissertation, we determined the conditions under which evacuees would be likely to trust a robot in an emergency evacuation. We began by examining reports of real-world evacuations and considering how guidance robots can best help. We performed two simulations of evacuations and learned that robots could be helpful as long as at least 30% of evacuees trusted their guidance instructions. We then developed several methods for a robot to communicate directional information to evacuees. After performing three rounds of evaluation using virtually, remotely and physically present robots, we concluded that robots should communicate directional information by gesturing with two arms. Next, we studied the effect of situational risk and the robot's previous performance on a participant's decision to use the robot during an interaction. We found that higher risk scenarios caused participants to align their self-reported trust with their decisions in a trust situation. We also discovered that trust in a robot drops after a single error when interaction occurs in a virtual environment. After an exploratory study in trust repair, we have learned that a robot can repair broken trust during the emergency by apologizing for its prior mistake or giving additional information relevant to the situation. Apologizing immediately after the error had no effect. Robots have the potential to save lives in emergency scenarios, but could have an equally disastrous effect if participants overtrust them. To explore this concept, we created a virtual environment of an office as well as a real-world simulation of an emergency evacuation. In both, participants interacted with a robot during a non-emergency phase to experience its behavior and then chose whether to follow the robot’s instructions during an emergency phase or not. In the virtual environment, the emergency was communicated through text, but in the real-world simulation, artificial smoke and fire alarms were used to increase the urgency of the situation. In our virtual environment, we confirmed our previous results that prior robot behavior affected whether participants would trust the robot or not. To our surprise, all participants followed the robot in the real-world simulation of an emergency, despite half observing the same robot perform poorly in a navigation guidance task just minutes before. We performed additional exploratory studies investigating different failure modes. Even when the robot pointed to a dark room with no discernible exit the majority of people did not choose to exit the way they entered. The conclusions of this dissertation are based on the results of fifteen experiments with a total of 2,168 participants (2,071 participants in virtual or remote studies conducted over the internet and 97 participants in physical studies on campus). We have found that most human evacuees will trust an emergency guidance robot that uses understandable information conveyance modalities and exhibits efficient guidance behavior in an evacuation scenario. In interactions with a virtual robot, this trust can be lost because of a single error made by the robot, but a similar effect was not found with real-world robots. This dissertation presents data indicating that victims in emergency situations may overtrust a robot, even when they have recently witnessed the robot malfunction. This work thus demonstrates concerns which are important to both the HRI and rescue robot communities.
TAR: Trajectory adaptation for recognition of robot tasks to improve teamwork

(Georgia Institute of Technology, 2015-11-10) Novitzky, Michael

One key to more effective cooperative interaction in a multi-robot team is the ability to understand the behavior and intent of other robots. Observed teammate action sequences can be learned to perform trajectory recognition which can be used to determine their current task. Previously, we have applied behavior histograms, hidden Markov models (HMMs), and conditional random fields (CRFs) to perform trajectory recognition as an approach to task monitoring in the absence of commu- nication. To demonstrate trajectory recognition of various autonomous vehicles, we used trajectory-based techniques for model generation and trajectory discrimination in experiments using actual data. In addition to recognition of trajectories, we in- troduced strategies, based on the honeybee’s waggle dance, in which cooperating autonomous teammates could leverage recognition during periods of communication loss. While the recognition methods were able to discriminate between the standard trajectories performed in a typical survey mission, there were inaccuracies and delays in identifying new trajectories after a transition had occurred. Inaccuracies in recog- nition lead to inefficiencies as cooperating teammates acted on incorrect data. We then introduce the Trajectory Adaptation for Recognition (TAR) framework which seeks to directly address difficulties in recognizing the trajectories of autonomous vehicles by modifying the trajectories they follow to perform them. Optimization techniques are used to modify the trajectories to increase the accuracy of recognition while also improving task objectives and maintaining vehicle dynamics. Experiments are performed which demonstrate that using trajectories optimized in this manner lead to improved recognition accuracy.
Physics-based reinforcement learning for autonomous manipulation

(Georgia Institute of Technology, 2015-08-21) Scholz, Jonathan

With recent research advances, the dream of bringing domestic robots into our everyday lives has become more plausible than ever. Domestic robotics has grown dramatically in the past decade, with applications ranging from house cleaning to food service to health care. To date, the majority of the planning and control machinery for these systems are carefully designed by human engineers. A large portion of this effort goes into selecting the appropriate models and control techniques for each application, and these skills take years to master. Relieving the burden on human experts is therefore a central challenge for bringing robot technology to the masses. This work addresses this challenge by introducing a physics engine as a model space for an autonomous robot, and defining procedures for enabling robots to decide when and how to learn these models. We also present an appropriate space of motor controllers for these models, and introduce ways to intelligently select when to use each controller based on the estimated model parameters. We integrate these components into a framework called Physics-Based Reinforcement Learning, which features a stochastic physics engine as the core model structure. Together these methods enable a robot to adapt to unfamiliar environments without human intervention. The central focus of this thesis is on fast online model learning for objects with under-specified dynamics. We develop our approach across a diverse range of domestic tasks, starting with a simple table-top manipulation task, followed by a mobile manipulation task involving a single utility cart, and finally an open-ended navigation task with multiple obstacles impeding robot progress. We also present simulation results illustrating the efficiency of our method compared to existing approaches in the learning literature.
Semantic mapping for service robots: building and using maps for mobile manipulators in semi-structured environments

(Georgia Institute of Technology, 2015-04-08) Trevor, Alexander J. B.

Although much progress has been made in the field of robotic mapping, many challenges remain including: efficient semantic segmentation using RGB-D sensors, map representations that include complex features (structures and objects), and interfaces for interactive annotation of maps. This thesis addresses how prior knowledge of semi-structured human environments can be leveraged to improve segmentation, mapping, and semantic annotation of maps. We present an organized connected component approach for segmenting RGB-D data into planes and clusters. These segments serve as input to our mapping approach that utilizes them as planar landmarks and object landmarks for Simultaneous Localization and Mapping (SLAM), providing necessary information for service robot tasks and improving data association and loop closure. These features are meaningful to humans, enabling annotation of mapped features to establish common ground and simplifying tasking. A modular, open-source software framework, the OmniMapper, is also presented that allows a number of different sensors and features to be combined to generate a combined map representation, and enabling easy addition of new feature types.
Timing multimodal turn-taking in human-robot cooperative activity

(Georgia Institute of Technology, 2015-04-07) Chao, Crystal

Turn-taking is a fundamental process that governs social interaction. When humans interact, they naturally take initiative and relinquish control to each other using verbal and nonverbal behavior in a coordinated manner. In contrast, existing approaches for controlling a robot's social behavior do not explicitly model turn-taking, resulting in interaction breakdowns that confuse or frustrate the human and detract from the dyad's cooperative goals. They also lack generality, relying on scripted behavior control that must be designed for each new domain. This thesis seeks to enable robots to cooperate fluently with humans by automatically controlling the timing of multimodal turn-taking. Based on our empirical studies of interaction phenomena, we develop a computational turn-taking model that accounts for multimodal information flow and resource usage in interaction. This model is implemented within a novel behavior generation architecture called CADENCE, the Control Architecture for the Dynamics of Embodied Natural Coordination and Engagement, that controls a robot's speech, gesture, gaze, and manipulation. CADENCE controls turn-taking using a timed Petri net (TPN) representation that integrates resource exchange, interruptible modality execution, and modeling of the human user. We demonstrate progressive developments of CADENCE through multiple domains of autonomous interaction encompassing situated dialogue and collaborative manipulation. We also iteratively evaluate improvements in the system using quantitative metrics of task success, fluency, and balance of control.