Search

Dynamically-balancing robots have recently been made available by Segway LLC, in the form of the Segway RMP (Robot Mobility Platform). We have addressed the challenge of using these RMP robots to play soccer, building up upon our extensive previous work in this multi-robot research domain. In this paper, we make three contributions. First, we present a new domain, called Segway Soccer, for investigating the coordination of dynamically formed, mixed human-robot teams within the realm of a team task that requires realtime decision making and response. Segway Soccer is a game of soccer between two teams consisting of both Segway riding humans and Segway RMPs. We believe Segway Soccer is the first game involving both humans and robots in cooperative roles and with similar capabilities. In conjunction with this new domain, we present our work towards developing a soccer playing robot using the RMP platform with vision as its primary sensor. Our third contribution is that of skill acquisition from a human teacher, where the learned skill is then used seamlessly during robot execution as part of its control hierarchy. Skill acquisition and use addresses the challenge of rapidly developing the low-level actions that are environment dependent and are not transferable across robots.

Recently, the company Segway LLC has released a dynamically balancing robot base, the Segway RMP, to complement their Segway scooters for human mobility. These robot bases provide exceptional robustness and capability with the unique feature of dynamic balancing at a human-size scale. We have addressed the challenge of using these Segway RMPs to build robots that are able to autonomously play soccer, building up upon our extensive previous work in this multi-robot research domain. This paper details our investigations towards developing an individually autonomous and capable robot. In particular, the focus is on the electro-mechanical mechanisms required to make a Segway RMP autonomous, able to sense its world, as well as manipulate and kick a soccer ball. In conjunction with the mechanisms required to make the robot physically capable, we detail our investigations into the control algorithms required to enable the robot to perceive, think, and act in real time for a dynamically changing world. While these techniques are applicable to many robot applications, dynamic balancing creates a number of unique challenges and opportunities that must be addressed. We examine these capabilities and limitations of the Segway and provide a detailed analysis of different mechanical and computational techniques to address these limitations. In this paper, we present empirical results examining the performance of our mechanisms and algorithms in the context of a fully functioning system

This thesis investigates the implementation of a biologically plausible navigation system on a mobile robot platform for the dual purpose of furthering robot navigation research and understanding of biological navigation systems. The aim of this work is to take steps towards bridging the broad gulf that exists between robot navigation research and biological navigation research and modelling. This dissertation reviews the extensive literature in three diverse fields of research; biological studies of animal navigation, computational modelling of animal navigation and robot navigation research. In particular, this thesis focuses on studies of rats with an emphasis on their cognitive navigation abilities. The hippocampus, which forms a large portion of the rat cerebral cortex, is believed to play a central role in the rat’s spatial navigation mechanisms and forms the primary focus of this thesis. Studies on the rat hippocampus, while the animals perform free spatial navigation tasks, have begun to reveal an increasingly detailed picture of how rats represent their location within the world. In particular, it appears that rats have separate representations that encode their spatial location and orientation in a world centred context independent of reward conditions. Most results indicate that these spatial representations are formed through the integration of ideothetic movement information, called path integration, with sensory stimuli. Indeed, it seems that the animals predominantly use ideothetic mechanisms to update their spatial representations and rely on sensory stimuli to correct for the inevitable error that accumulates. Numerous models of rat navigation, based on the observed biological results, have been proposed with varying degrees of simulation. In general, no model is complete, or comprehensive, in the sense that it addresses all the implementation issues in a systematic and realistic manner. In contrast, successful robot navigation systems, by necessity, do address all the implementation issues. This thesis represents an attempt at implementing a rat navigation model on a real robot platform with an eye for developing an autonomous system. The VIPER robot was custom built for this task and represents a semipowerful, general purpose, robot platform with monocular vision as the primary sensor. This thesis describes and validates, the navigation architecture developed for the VIPER robot. The system takes a holistic approach with the underlying guide of biological plausibility and autonomous operation. This thesis reports on the investigations into reactive navigation and high-speed reactive navigation along with investigations into place and goal memory systems. In all cases, the investigations focused on pragmatic biologically plausibility. That is, biological plausibility without compromising the computational performance, or autonomy of the system. Hence, the robot operates in an autonomous manner in real-time for all the investigations reported in this dissertation. Many of the robot investigations proved highly successful, however, due to time constraints it was not possible to achieve the full aim of the project and make the system completely autonomous. Moreover, some of the system components proved only semi-reliable. Where the navigation system is incomplete, or not totally successful, preliminary investigations of better mechanisms were carried and are reported in this dissertation. Thus, while this thesis in itself does not bridge the gap between biology and robotics it represents a significant step towards that direction.

In this paper, we focus on human-robot interaction in a team task where we identify the need for peer-to-peer (P2P) teamwork, with no fixed hierarchy for decision making be- tween robots and humans. Instead, all team members are equal participants and decision making is truly distributed. We have fully developed a P2P team within Segway Soccer, a research domain, built upon Robocup robot soccer, that we have introduced to explore the challenge of P2P coor- dination in human-robot teams with dynamic, adversarial tasks. We recently participated in the first Segway Soccer games between two competing teams at the 2005 RoboCup US Open. We believe these games are the first ever between two human-robot P2P teams. Based on the competition, we realized two different approaches to P2P teams. We present our robot-centric approach to P2P team coordination and contrast it to the human-centric approach of the opponent team.

As robots become more complex by incorporating dynamic stability or greater mechanical degrees of freedom, the difficulty of developing control algorithms directly on the robot increases. This is especially true for large or expensive robots, where damage is costly, or where communication bandwidth limits in-depth debugging. One effective solution to this problem is the use of a flexible, physically-accurate simulation environment which allows for experimentation with the physical composition and control systems of one or more robots in a controlled virtual setting. While many robot simulation environments are available today, we find that achieving accurate simulation of complex, vision-centric platforms such as the Segway RMP or Sony AIBO requires accurate modeling of latency and robust synchronization. Building on our previous work, we present an open-source simulation framework, ÜberSim, and demonstrate its ability to simulate vision-centric, balancing robots in a realistic fashion. The focus of this simulation environment is on accurate simulation with high-frequency control loops and flexible configuration of robot structure and parameters via a client-side definition language.

Coordinated action for a team of robots is a challenging problem, especially in dynamic, unpredictable environments. In the context of robot soccer, a complex domain with teams of robots in an adversarial setting, there is a great deal of uncertainty in the opponent’s behavior and capabilities. We introduce the concept of a play as a team plan, which combines both reactive principles, which are the focus of traditional approaches for coordinating actions, and deliberative principles. We introduce the concept of a playbook as a method for seamlessly combining multiple team plans. The playbook provides a set of alternative team behaviors which form the basis for our third contribution of play adaptation. We describe how these concepts were concretely implemented in the CMDragons robot soccer team. We also show empirical results indicating the importance of adaptation in adversarial or other unpredictable environments

"Recently, we have begun investigating a new robot soccer domain built around the concept of human-robot teams in a 'peer' setting. One of the key challenges for addressing effective human-robot interaction is to robustly identify and track people and robot teammates without requiring undue prior knowledge of their appearance. For cost and complexity reasons, our robots are equipped with monocular color cameras. Thus, we seek an algorithm to enable reliable acquisition and tracking of people and robots from a robot armed with a monocular color camera. We have developed a novel algorithm for acquiring and tracking a single human subject from a dynamically balancing platform, a Segway RMP robot, using a monocular color camera. Our technique uses a combination of known vision and tracking techniques including region growing, motion detection, and mean-shift color-template tracking. In this paper, we describe our approach, and analyze its performance and limitations, for both acquiring and tracking a single human target in an indoor environment. Our experiments demonstrate that acquisition and tracking are feasible with a monocular camera even for a dynamically balancing platform. Moreover, our results show that with current processor technology real-time tracking and robot response are available."

In this work we develop a visual place recognition system for omnidirectional panoramic images that makes use of their full 360 degree circumferential angle of view. The system builds on our previous variant of the sequence SLAM algorithm to include panoramic image rotation alignment for improved image similarity scoring, and bi-directional query and database sequence matching within a Hidden Markov Model (HMM) framework for robust place recognition. Three rotation alignment methods are explored including image pixel registration, and both image cross correlation and zero phase alignment in the spherical Fourier domain. All alignment methods operate using low-resolution images for computational efficiency. Experiments using an outdoor panoramic image dataset demonstrate improved precision recall performance using rotation alignment and bi-directional sequence matching. In particular, place recognition is possible in scenarios where a robot traverses a previous path in the opposite direction.

Actuated lidar, where a scanning lidar is combined with an actuation mechanism to scan a 3D volume rather than a single line, has been used heavily on a wide variety of field robotics applications. Common examples of actuated lidar include spinning/rolling and nodding/pitching configurations. Due to construction of actuated lidar, the center of rotation of the lidar mirror may not coincide with the center of rotation of the actuation mechanism. In order to triangulate a precise point cloud representation of the environment, the centers of rotation must be brought into alignment using a suitable calibration procedure. We refer to this problem as estimating the internal parameters of actuated lidar. In this work, we focus on spinning/rolling lidar and present a fully automated algorithm for calibration using generic scenes without the need for specialized calibration targets. The algorithm is evaluated on a range of real and synthetic data and is shown to be robust, accurate and has a large basin of convergence.

We present a direct visual odometry formulation using a warping function in disparity space. In disparity space measurement noise is well-modeled by a Gaussian distribution, in contrast to the heteroscedastic noise in 3D space. In addition, the Jacobian of the warp separates the rotation and translation terms, enabling motion to be estimated from all image points even those located at infinity. Furthermore, we show that direct camera tracking can obtain accurate and robust performance using only a fraction of the image pixels through a simple and efficient pixel selection strategy. Our approach allows faster than real-time computation on a single CPU core with unoptimized code. As our approach does not rely on feature extraction, the selected pixels over successive frames are often unique. Hence, triangulating the selected pixels to the world frame produces an accurate and dense 3D reconstruction with minimal computational cost making it appealing to robotics and embedded applications. We evaluate the performance of our approach against state-of-the-art methods on a range of urban and indoor datasets. We show that our algorithm produces competitive performance, requires no specialized tuning, and continues to produce competitive results even when run with low resolution images where other techniques fail to operate.

We present a vision-based mapping and localization system for operations in pipes such as those found in Liquified Natural Gas (LNG) production. A forward facing, fisheye camera mounted on a prototype robot collects imagery as it is tele-operated through a pipe network. The images are processed offline to estimate camera pose and sparse scene structure where the results can be used to generate 3D renderings of the pipe surface. The method extends state of the art visual odometry and mapping for fisheye systems to incorporate geometric constraints based on prior knowledge of the pipe components into a Sparse Bundle Adjustment framework. These constraints significantly reduce inaccuracies resulting from the limited spatial resolution of the fisheye imagery, limited image texture, and visual aliasing. Preliminary results are presented for a dataset collected in fiberglass pipe network which demonstrate the validity of the approach.

Robust motion control algorithms are fundamental to the successful, autonomous operation of mobile robots. Motion control is known to be a difficult problem, and is often dictated by a policy, or state-action mapping. In this chapter, we present an approach for the refinement of mobile robot motion control policies, that incorporates corrective feedback from a human teacher. The target application domain of this work is the low-level motion control of a mobile robot. Within such domains, the rapid sampling rate and continuous action space of policies are both key challenges to providing policy corrections. To address these challenges, we contribute advice-operators as a corrective feedback form suitable for providing continuous-valued corrections, and Focused Feedback For Mobile Robot Policies (F3MRP) as a framework suitable for providing feedback on policies sampled at a high frequency. Under our approach, policies refined through teacher feedback are initially derived using Learning from Demonstration (LfD) techniques, which generalize a policy from example task executions by a teacher.We apply our techniques within the Advice-Operator Policy Improvement (A-OPI) algorithm, validated on a Segway RMP robot within a motion control domain. A-OPI refines LfD policies by correcting policy performance via advice-operators and F3MRP. Within our validation domain, policy performance is found to improve with corrective teacher feedback, and moreover to be similar or superior to that of policies provided with more teacher demonstrations.

Structure-From-Motion (SFM) methods, using stereo data, are among the best performing algorithms for motion estimation from video imagery, or visual odometry. Critical to the success of SFM methods is the quality of the initial pose estimation algorithm from feature correspondences. In this work, we evaluate the performance of pose estimation algorithms commonly used in SFM visual odometry. We consider two classes of techniques to develop the initial pose estimate: Absolute Orientation (AO) methods, and Perspective-n-Point (PnP) methods. To date, there has not been a comparative study of their performance on robot visual odometry tasks. We undertake such a study to measure the accuracy, repeatability, and robustness of these techniques for vehicles moving in indoor environments and in outdoor suburban roadways. Our results show that PnP methods outperform AO methods, with P3P being the best performing algorithm. This is particularly true when stereo triangulation uncertainty is high due to a wide Field of View lens and small stereo-rig baseline.