Here are some of the demos and posters that were shown:

Southern California Ground Model [PDF]

Generating detailed, high-resolution models of the physical characteristics of the Earth is the first step in earthquake simulation. Ground model datasets are expensive to generate, store, validate, and compare. They are usually generated and managed in an ad-hoc fashion. Our goal is to use clusters to automatically generate, store, and manage ground model datasets, in a form that can be easily validated and shared among groups of seismologists. Our ground model generator runs on a cluster of blade servers and uses the Hadoop runtime system, an open-source implementation of Google's Map/Reduce parallel programming model. The demo will show the generator running live on the cluster, as well as visualizations of generated velocity models.

Authors:

Optimizing Data Transfers through Intelligent Resource Scheduling [PDF]

This demo presents dsync, a new file transfer system, that intelligently uses available resources to reduce transfer times. Dsync can download from the original sender, peers, and also search the local disk (using a set of heuristics that help identify likely similar files). The system does not make any assumptions about the environment (network speed, disk load, etc.) and dynamically adapts throughout the transfer. The demo shows how dsync's novel optimization framework can synchronize a collection of nodes faster than traditional tools such as BitTorrent and rsync.

Authors: Bindu Pucha (Purdue University), David Andersen (Carnegie Mellon University), Michael Kaminsky, Michael Kozuch

Introduction to Dynamic Physical Rendering (DPR) [PDF]

Intel and Carnegie Mellon are collaborating on a new form of programmable matter that can create moving, physical, 3D objects by using thousands to millions of tiny, robotic modules. Potential applications of this research include medicine, product design, human communications (teleconferencing, telepresence), smart antennas, 3D faxing, and general purpose robots adaptable to any task.
Central to the DPR effort is the need to develop more approaches to programming, coordination, and debugging for systems involving very high levels of concurrency (tens of thousands to millions of autonomous, interacting nodes). In the near term, such new models for control and fault isolation could benefit many other classes of highly parallel and dynamic computing systems.

Meld: Declarative Ensemble Programming [PDF]

We show Meld, a new language for modular robot programming inspired by declarative languages in other domains including the P2 language for defining overlay networks and routing. Programming DPR ensembles presents many special challenges due to the constant rearrangement of the ensemble, the vagaries of moving parts, and the potential for temporary or permanent faults. We believe that declarative programming can bring the same level of expressiveness, conciseness, and global view to programming robotic ensembles as P2 has for programming overlay networks. We show how Meld programs are structured and executed, illustrating several examples in the DPR simulator.

Authors: Michael Ashley Rollman (IRP Intern/CMU), Babu Pillai, Sidd Srinivasa, Jason Campbell, Seth Copen Goldstein (CMU), Todd Mowry, Peter Lee (CMU)

LDP: Programming with Distributed Conditions [PDF]

We demonstrate LDP, a reactive programming language that uses locally-distributed predicates as triggers for the execution of specified actions. This effort builds on work we did last year using similar distributed condition triggers for debugging. In contrast to that older work, we extend the condition language to control actions and add new operators suited to more general purpose use. In this poster and demo we present the overall structure of the programming language, illustrate how the distributed condition detection and execution mechanism works, and demonstrate the generality and usefulness of this programming approach using several example programs running in DPRSim.

Authors: Michael De Rosa (IRP Intern/CMU), Jason Campbell, Babu Pillai, Seth Copen Goldstein (CMU), Peter Lee (CMU), Sidd Srinivasa, Todd Mowry

Cubic Electrostatic Adhesion Prototypes for Catoms and Modular Robots [PDF]

We demonstrate a new modular robot design we have developed based on electrostatic actuator plates as a reliable and switchable adhesion mechanism between adjacent modules. Our design roughly follows several prior cubic modular robot systems in that the faces of our cubes telescope to permit an ensemble to reconfigure itself. But unlike prior cubic modular robots we apply electric-fields (or more specifically, electric-field--modulated frictional forces) to hold neighboring modules together. This allows us to simplify the module design, and ultimately will allow catoms to be.

Authors: Michael Weller (CMU), Brian Kirby (CMU), Emre Karagozler (CMU), Seth Copen Goldstein (CMU), Jason Campbell

Electromagnetic Catom Prototypes and DPRsim Integration [PDF]

We show new prototype catom modules which, in contrast to prototypes shown previously, are sufficiently powerful and sophisticated to work reliably in groups of three and more catoms - a significant milestone in our hardware research.. These illustrate successful integration of computation, communications, and actuation, the three basic ingredients required for any catom. Although constructed at 5cm scale to ease manufacturing and assembly, and although constrained to a 2D plane, they allow us to experiment with the local coordination techniques required to generate controlled motion with distributed electromagnetic or electrostatic actuators. An important secondary aspect of these prototype is the absence of moving parts - an attribute which may be very important for high-volume manufacturing.

Authors: Brian Kirby (CMU), Seth Copen Goldstein (CMU) , Burak Aksak (CMU), Jason Campbell, Babu Pillai

Millimeter-scale Electrostatic Actuation Prototypes [PDF]

Our scaling analysis of DPR module designs indicates that electric field actuation is viable at millimeter-scale and smaller sizes. To help substantiate this hypothesis, we have developed this hardware demonstration of electrostatic actuation of a small metal cylinder. This illustrates not only the controlled motion of a millimeter-scale object using electric fields, but also actuation between an active and a passive object, using the effects of a mirror charge developed in the passive surface.

Authors: Emre Karagozler, Jason Campbell, Babu Pillai, Seth Copen Goldstein (CMU), Gary Fedder (CMU)

Internal Localization in Large Catom Ensembles [PDF]

A key assumption in much of the ongoing work in DPR is that an ensemble can estimate the position of its members - a problem we call "internal localization". This problem is straightforward given perfect measurements of the angles and distances between all the members, but grows very challenging with any realistic sensor model. Uncertain observations, absence of external references, and the large number of modules involved make many approaches intractable (and even severely limited the usefulness of several prior approaches we developed). This exhibit illustrates a new, scalable, hierarchical approach to internal localization, which we show can achieve high localization accuracy very quickly in variety of 2D and 3D configurations.

Authors: Stano Funiak (CMU), Babu Pillai, Jason Campbell, Seth Copen Goldstein (CMU), Carlos Guestrin (CMU)

A Planning Framework for Modular Robots [PDF]

Part I - Redesigning metamorphic systems for easy reconfiguration: Reconfiguration planning in large-scale metamorphic systems is complicated by the combinatorial explosion of the state space, necessitating good heuristics. Generating good heuristics is an ongoing area of research; one major difficulty is the existence of blocking constraints --- in a tightly packed ensemble, motion of the modules in the interior is blocked by their neighbors. We describe a system that makes reconfiguration planning significantly easier. We do this by characterizing the nonholonomic constraints common in metamorphic systems and designing a metamodule-based system that avoids them. Freeing the system of blocking constraints allows us to plan fast, provably complete plans for many thousands of modules.
Part II - A provably complete distributed planner for shape change: This demo demonstrates a distributed planning algorithm for large-scale shape change in metamorphic systems. A distributed algorithm running on each module is able to affect shape change while maintaining connectivity of the entire ensemble. The algorithm has been implemented in a new declarative programming language, MELD, created specifically for modular robots. Programming in a formal language like MELD allows us to use automated theorem provers, like TWELF, to prove the completeness of the planner.

Authors: Daniel Dewey (CMU), Michael Ashley-Rollman (IRP Intern/CMU), Siddhartha Srinivasa

Speculative Packet Forwarding in DPR [PDF]

DPR ensembles can be thought of as very large communication networks, with only geographically local links. The resulting topology can have many thousands of hops between arbitrary nodes, and will suffer from high average communication latencies. This poster presents a novel approach to reducing latencies by generating an ad-hoc overlay network employing fast cut-through routing with speculation for long, multi-hop overlay links. The result is a system that can realize cross-ensemble communication with just a fraction of the latency possible with traditional routing and forwarding.

Authors: Ben Rister, Babu Pillai, Phil Gibbons, Todd Mowry

Cell Tracking [PDF]

Automated visual-tracking of cell populations in vitro using phase contrast time-lapse microscopy is vital for quantitative, systematic and high-throughput measurements of cell behaviors. The low signal-to-noise ratio of phase contrast microscopy images, high and varying densities of the cell cultures, topological complexities of cell shapes, and the wide range of cell behaviors pose many challenges to existing tracking techniques. We present a fully-automated multi-target tracking system that can efficiently cope with these challenges, and can simultaneously track and analyze thousands of cells. The system is designed for online tracking during image acquisition. We applied this methodology to tracking a range of cells including adult stem cells. Tracking accuracies in the range of 84%-92% were achieved on large-scale experiments. This tracking methodology has broad applications in tissue engineering, stem cell research, drug discovery and development, and related areas.

Authors: Kang Li (CMU), Mei Chen, Takeo Kanade (CMU)

ISADS for Skin Cancer [PDF]

Skin cancer is the most common cancer in the U.S., and melanoma accounts for more than 75% of skin cancer deaths. However, despite its fatality, if discovered early, skin cancer is highly curable. Dermoscopy is a non-invasive imaging modality for skin cancer screening, a first line defense. Traditionally doctors augment personal expertise with medical literature that is typically indexed by disease rather than by relevance to the current case. We demonstrate a system that enables doctors to make more informed decisions about a given case by presenting relevant annotated cases from large medical repositories, where the retrieval is based on image content (not text/metadata). The underlining technologies are medically-relevant feature extraction and classification, and similarity computing in high-dimensional feature spaces.

Authors: Mei Chen, Richard Gass, Howard Zhou (Georgia Tech), Laura Ferris (UPMC), Jon Ho (UPMC), Laura Drogowski (UPMC & Pitt)

Just-in-Time Indexing for Interactive Data Exploration [PDF]

Interactive search of complex data poses significant challenges for traditional indexing methods because of the infeasibility of determining an effective set of indices a priori. We propose just-in-time indexing, a new strategy that mitigates these challenges by exploiting a key characteristic of interactive data exploration: iterative query refinement. During the refinement process, just-in-time indexing takes advantage of user think time to create indices on-the-fly for query terms likely to be relevant to the current user. Moreover, because a user typically refines a query after observing only a subset of the results, just-in-time indexing indexes only subsets of the data at a time. Our poster describes just-in-time indexing, details strategies for balancing the needs of the current user (immediate workload) versus the projected needs of future users (long-term workload), and reports on an experimental study of users performing photo searches on our OpenDiamond platform with and without just-in-time indexing.

Authors: Lily Mummert, Phil Gibbons, Rahul Sukthankar, M. Satyanarayanan (CMU)

ISADS for Breast Cancer [PDF]

The goal of ISADS is to enable doctors to make better decisions about a given case by providing a selection of similar annotated cases from a large database. This demo shows ISADS applied to mammograms. A mass region of interest (ROI) is identified, characterized as a set of features, and then compared against a database of ROIs. The most similar cases are shown with outcomes. A fundamental challenge in developing ISAD systems is the identification of similar cases, not simply in terms of superficial image characteristics, but in a medically-relevant sense. We highlight algorithms for learning distance metrics with emphasis on visual similarity.

Authors: Lily Mummert, Rahul Sukthankar (Intel Research Pittsburgh) Bin Zheng (UPMC/U.Pitt) Rong Jin, Liu Yang (MSU) Adam Goode, Satya (CMU)

OCT Quality Classification [PDF]

Optical Coherence Tomography (OCT) is an imaging technique that allows the retina to be imaged with micrometer precision. However, OCT image quality varies both globally in the entire region, as well as locally in parts of the image (due to eye movement, pathology, or physiology). Current evaluation techniques give global quality scores, but often there are local distortions that can improperly raise or lower an image's overall quality score. We present a machine learning based technique that hierarchically classifies OCT image quality using medically relevant image features. The experiment on 270 images is cross-validated by three domain experts.

Authors: Peter Barnum (CMU), Mei Chen, Hiroshi Ishikawa (UPMC & Pitt), Gadi Wollstein (UPMC & Pitt), Joel Schuman (UPMC & Pitt)

3-D Segmentation [PDF]

We present a novel method for 3D medical image segmentation that is fully automated following a simple initialization procedure of selecting a 2D slice and tracing a 2D target within. Our algorithm performs n-dimensional segmentation by placing spheres of independently variable radius centered at each image voxel, which are grown to reach, but not cross, the nearest object boundary. Variable Scale Statistics (VSS) are computed on the voxel populations within spheres as they grow to guide their evolution. We customize each 3D segmentation to the specific data set and target object by automatically calculating optimal 3D algorithm parameters from the initial 2D tracing. The quick and simple initialization procedure requires no direct manual parameter initialization, allowing clinical users unfamiliar with technical specifics to effectively segment 3D objects of complex shapes.

Authors: Aaron Cois (Pitt), Mei Chen, George Stetten (CMU & Pitt)

Storage Systems for Discard-Based Search [PDF]

Interactive search of complex data presents a distinctive workload to the storage system. Data objects tend to be large, and are read in their entirety. Objects may be delivered in any order, allowing the storage system the flexibility to determine the most efficient order. Objects are processed independently, providing the potential for a high degree of concurrency in both processing and storage, even within a single search. We examine some implications of these workload characteristics on storage system design. For example, we show that a JBOD ("just a bunch of disks") organization is more effective than striping, and that any-order delivery enables several optimizations, the most important of which is an I/O coalescing technique called "bandwagon synchronizationr".

Authors: Lily Mummert, Steve Schlosser, Mike Mesnier, M. Satyanarayanan (CMU)

Interactive Detection of Anomalies in Large Collections of Digital Microscopy Images [PDF]

The advent of high-throughput microscopy has enabled the creation of large collections of cell images for applications in drug discovery. However, as the size of these databases grows, it becomes increasingly important to focus the human scientists' limited time on those images of greatest value. Typically, the user would like to examine those images that are different from the rest, or outliers. However, the definition of an outlier is highly application-specific, and motivates the need for an interactive system for outlier detection (e.g., "show me images that contain an abnormally small number of cells"). We demonstrate a Diamond application for interactive outlier detection that distributes the search over a cluster of storage devices and computes the outlier statistics in a completely online and parallel manner. This enables our system to efficiently find outliers in a single pass through the data.

Authors: Adam Goode, Satya, (CMU) David Ross, Anil Tarachandani, Jeff Saltzman, (Merck) Mei Chen, Lily Mummert, Rahul Sukthankar (Intel Research)

Safer Software Execution through LBA [PDF]

Lifeguards are software tools that proactively monitor running computer programs and watch for potential problems such as virus attacks. Log-Based Architectures (LBA) are a new class of potential hardware enhancements designed to improve the performance of fine-grained lifeguards running on multicore platforms. To demonstrate the power of LBA, we will show a lifeguard using such enhancements in a multicore full-system simulator to detect a computer virus attack and prevent the virus from causing real damage to the system. Specifically, we show an attack that successfully steals user credit card numbers but is thwarted once the LBA lifeguard is activated.

Authors: Shimin Chen, Phil Gibbons, Mike Kozuch, Todd Mowry, Michael Ryan, Natassa Ailamaki (CMU), Babak Falsafi (CMU), Tunji Ruwase (CMU), Theodore Strigkos (CMU), Evangelos Vlachos (CMU)

Monitoring Parallel Applications using LBA [PDF]

Parallelism, particularly multicore-style parallelism, plays two important roles in Log-Based Architectures. First, multicore processors enable lifeguards to execute in parallel with the applications they monitor. Second, LBA is designed to enable lifeguards that efficiently monitor parallel applications. This demo will highlight this second role by showing a multithreaded lifeguard efficiently monitoring a multithreaded application for potential data races. This addresses a key challenge in parallel programming: avoiding unwanted data races.

Authors: Shimin Chen, Phil Gibbons, Mike Kozuch, Todd Mowry, Michael Ryan, Natassa Ailamaki (CMU), Babak Falsafi (CMU), Tunji Ruwase (CMU), Theodore Strigkos (CMU), Evangelos Vlachos (CMU)

Parallelizing Sophisticated Lifeguards [PDF]

Sophisticated Lifeguards perform considerable work for each program event logged by LBA. A complementary approach to aggressive event filtering (as discussed above for the LBA Event Filtering Demo) is to parallelize the lifeguard functionality across multiple cores. However, important lifeguards can have significant serial dependencies that make them challenging to parallelize. This demo animates our successful parallelization of one such lifeguard. In this scheme, segments of the log are partitioned among the cores for parallel processing. We also present new results on parallel-friendly compression of log segments.

Authors: Shimin Chen, Phil Gibbons, Mike Kozuch, Todd Mowry, Michael Ryan, Natassa Ailamaki (CMU), Babak Falsafi (CMU), Tunji Ruwase (CMU), Theodore Strigkos (CMU), Evangelos Vlachos (CMU)

Improving LBA Performance through Event Filtering [PDF]

While monitoring the execution of an application, a fine-grained software lifeguard inspects an enormous number of program events every second (typically a billion or more). Thus, eliminating unnecessary events is a key performance optimization in Log-Based Architecture systems. This demo will show a comparison between lifeguard performance with and without event filtering using two of the novel filters developed by the LBA team. The demo will also provide a good opportunity for viewing the information maintained in the LBA execution logs.

Authors: Shimin Chen, Phil Gibbons, Mike Kozuch, Todd Mowry, Michael Ryan, Natassa Ailamaki (CMU), Babak Falsafi (CMU), Tunji Ruwase (CMU), Theodore Strigkos (CMU), Evangelos Vlachos (CMU)

Future Optical Chip-to-Chip Interconnection Fabrics [PDF]

As device feature size scaling continues, interconnects become a key performance limiting factor in data transmission rates. This demo describes state of the art and a systems perspective to identify future technology gaps and challenges to chip-to-chip interconnect systems, including the multi-core architectures.
One specific challenge is an ultra-low power light sources (with threshold currents < 10 µA) that can be monolithically integrated into silicon chips. There is a new generation of light sources, including, nanowire light sources that might change this situation. Together with silicon-based modulators one might construct interconnect fabrics that easily integrate into silicon-based chips.
The long-term goal of this project is the exploration and development of a complete prototype, based on specific system level applications, including the active and passive components.

Authors: Madeleine Glick, Matt Chabalko (CMU), Elias Towe (CMU)

Polymer Waveguides for Low Cost Optical Backplanes [PDF]

Data rates in the backplane are increasing to several Gbps/channel and higher. This demo presents the use of polymer backplanes enabling parallel waveguide connections between cards. The polymer waveguides are bit rate transparent. With parallel optics this would enable a high capacity yet passive backplane. This switchless design should have a relatively lower initial cost for the customer. The Centre for Photonic Systems at the University of Cambridge and Dow Corning have previously developed novel polymer devices that can be integrated with printed circuit boards. Together we are exploring their potential for use in high-speed on-board optical networks. Recent results show efficient high-speed data transmission at 10 Gbps with low loss and excellent crosstalk performance.

Authors: Madeleine Glick (IRP), Joseph Beals IV (Cambridge University Engineering Department), Nikos Bamiedakis (Cambridge University Engineering Department), Adrian Wonfor (Cambridge University Engineering Department), Richard V. Penty (Cambridge University Engineering Department), Ian H. White (Cambridge University Engineering Department), Jon V. DeGroot Jr. (Dow Corning), Terry V. Clapp (Dow Corning)

Digital Signal Processing for Low Cost Optical Links [PDF]

Digital signal processing has the potential to overcome cost and power challenges associated with optical links. In particular, CMOS-based digital equalization techniques, recently developed for long-haul 10 Gbit/s communications, offer the greatest advantages in terms of functionality and reproducibility.
This demo describes our simulation and experimental investigations of the following fundamental issues:

• What digital processor architectures are most effective for this application?
• What improvements in the performance of optical communication links are possible using digital signal processing.

By answering these questions, the work will identify the impairments in short optical links that can be mitigated through the use of DSP, will develop effective digital filter architectures to implement this, and will quantify the cost savings and gains in system performance.

Authors: Madeleine Glick, Philip Watts (University College London), Robert Waegemans (University College London), Ramanan Thiruneelakandan (University College London), and Robert Killey (University College London)

Optimizing Both Private and Shared Cache Performance in Multicores [PDF]

In the near future, multicore processors will have tens to hundreds of processing cores. Good performance will require making good use of both the cores and the on-chip caches. We envision a setting in which fine-grained program tasks are automatically and dynamically scheduled on the cores, in order to achieve such good performance without burdening the programmer. In this poster, we report on a new online thread scheduling algorithm that has provably good core utilization and cache performance for a broad class of divide-and-conquer-style applications. Unlike previous work with provable cache performance, the analysis considers the combined and sometimes competing effects of having both exclusive caches and shared caches on-chip. We also present new results for cache-efficient algorithms for multiplying a Sparse Matrix by a Dense Vector, an important kernel for many applications. Finally, we present a trace-based approach for automatically selecting task granularities in order to optimize cache performance.

Authors: Phil Gibbons, Shimin Chen, Mike Kozuch, Guy Blelloch (CMU), Rezaul Chowdhury (UT Austin), Vijaya Ramachandran (UT Austin)

Insight: User-friendly Storage Management for the Digital Home [PDF]

A significant barrier to the increased use of digital data storage in the home is that it is too difficult for users to manage. We believe that manageability is like security - it is extremely important but difficult, if not impossible, to add to a system after the fact. Therefore, we have designed and are implementing Perspective, a storage system for the digital home built from the ground up with manageability as a principal design feature. Building on Perspective, Insight is a new set of storage management tools that users can employ to configure, monitor, and troubleshoot ensembles of Perspective storage devices. Insight and Perspective use semantic queries as a management primitive, enabling a wide variety of management tasks to be more intuitive to everyday users. This demo will allow users to use Insight to manage a set of home storage devices (e.g., desktop and laptop computers, music players, and digital video recorder).

Authors:

Event Detection in Crowded Videos [PDF]

Real-world actions occur often in crowded, dynamic environments. This poses a difficult challenge for current approaches to video event detection because it is difficult to segment the actor from the background due to distracting motion from other objects in the scene. We propose a technique for event recognition in crowded videos that reliably identifies actions in the presence of partial occlusion and background clutter. Our approach is based on three key ideas: (1) we efficiently match the volumetric representation of an event against oversegmented spatio-temporal video volumes; (2) we augment our shape-based features using flow; (3) rather than treating an event template as an atomic entity, we separately match by parts (both in space and time), enabling robustness against occlusions and actor variability. Our experiments on human actions, such as picking up a dropped object or waving in a crowd show reliable detection with few false positives.

Yan Ke (CMU), Rahul Sukthankar, Martial Hebert (CMU)

Introduction to Personal Robotics: Freeing Robots from the Factory Floor [PDF]

Most commercial robots function in "non-natural" environments, such as factories, where uncertainty has been engineered away. We believe that advances in sensing, machine learning, planning, and computing can enable robots to function usefully in less structured environments. The goal of this project is to enable robots to robustly perform useful tasks in uncertain environments such as the home, where it is infeasible to obtain complete information and the surroundings are constantly being modified.

Authors: Dave Ferguson, Ali Rahimi, Joshua Smith, Siddhartha Srinivasa, Casey Helfrich

The Robotic Barkeeper [PDF]

Developing general purpose robotic assistants that are able to manipulate our common objects and perform tasks in our environments has motivated and challenged robotic researchers for decades. However, recent advances in several key areas have brought us within reach of this goal. At Intel Research we are combining these developments to create a useful robot for homes and offices. This demonstration highlights our autonomous robotic arm and mobile robot and their ability to gather and manipulate common objects such as coffee mugs.

Authors: Dmitry Berenson, Rosen Diankov, Mike Vande Weghe (CMU), Dave Ferguson, Casey Helfrich, Siddhartha Srinivasa

Imitation Learning for Grasping [PDF]

Grasping complex objects with complex robot hands is an ongoing topic of research. However, humans are able to grasp complex objects and manipulate them easily. Imitation learning exploits the fact that it is much easier for a human to demonstrate a good grasp than to describe the complex nonlinear function being optimized to select that particular grasp. Our new functional gradient imitation learning algorithm learns good grasp evaluation metrics from demonstration. We demonstrate its effectiveness on using a Barret hand to grasp objects from the Princeton Shape Database, which comprises of complex real-world shapes made from many thousands of polygons.

Authors: Nathan Ratliff (CMU), Siddhartha Srinivasa, J. Andrew Bagnell (CMU)

Improving Sampling-based Planning [PDF]

Underspecified goals occur naturally in arm motion planning, where the final end-effector position (where we want the hand to be) is crucial but the configuration of the rest of the arm (where each individual joint should be) is not. Further, for complex mechanisms it is often impossible to uniquely specify the joint positions given a desired end-effector position. However, current planning approaches typically require a complete goal configuration to plan towards. In this work, we present a sampling-based planning algorithm capable of efficiently generating solutions for high-dimensional manipulation problems involving underspecified goals and complex, obstacle-laden environments. Our approach, known as Jacobian Transpose-directed Rapidly-exploring Random Trees, is able to combine the joint space exploration of the Rapidly-exploring Random Trees planning algorithm with a workspace goal bias to produce direct paths through complex environments extremely efficiently, without the need for inverse kinematics. We highlight results from both simulation and a physical robotic arm.

Authors: Mike Vande Weghe (CMU), Dave Ferguson, Siddhartha Srinivasa

openRave: A Planning Testbed for Robotic Systems [PDF]

Planning in complex scenes and managing all robot information is becoming more and more critical for today's robotic applications. We introduce a new open-source robot planning architecture called OpenRAVE that servers as the center of all high-level processing the robot has to perform in order to complete its task. Users of OpenRAVE only have to concentrate on the planning and scripting aspect of their task without having to worry about the complex details of collision detections, robot kinematics, dynamic world updates, robot controls, and scripting environments. The architecture is a plug-in based so that any planner, robot controller, and robot can be dynamically loaded without having to recompile the core code. This can enable the robotics community to easily share and compare algorithms with ease. OpenRAVE also supports a powerful network scripting environment which makes it easy to control robots and change execution flow during runtime of the task all using a program like Matlab.

Authors: Rosen Diankov, Dmitry Berenson, James Kuffner, Mike Vande Weghe (CMU) Dave Ferguson, Siddhartha Srinivasa

Grasp Planning in Complex Environments [PDF]

This project shows grasp analysis and manipulation planning techniques to perform fast grasp planning in complex scenes. In previous work on grasping, the object being grasped is assumed to be the only object in the environment. Hence the grasp quality metrics and grasping strategies developed do not perform well when the object is close to obstacles and many good grasps are infeasible. We introduce a framework for finding valid grasps in cluttered environments that combines a grasp quality metric for the object along with information about the local environment around the object. We encode these factors in a grasp-scoring function which we use to rank a precomputed set of grasps in terms of their appropriateness for a given scene. We show that this ranking is essential for efficient grasp selection and present experiments in simulation and on the Barrett/WAM robot.

Authors: Rosen Diankov, Dmitry Berenson, James Kuffner (CMU), Dave Ferguson, Siddhartha Srinivasa

Simultaneous Tracking, Segmentation, and Recognition of Objects [PDF]

The traditional computer vision paradigm involves three tasks: 1) segmentation, which partitions the 2-D space of an image into regions, and can be thought of as determining a region of interest, 2) tracking, which updates this region of interest through time, and 3) recognition, which provides a semantic label to the tracked entity. Traditionally, these three tasks have been integrated in a cascaded architecture. Unfortunately, while this independent processing leads to computationally-efficient implementations, the architecture also propagates errors from earlier stages to later stages in an unrecoverable manner. This observation has motivated recent work in attempting to integrate the three tasks more closely so that each task is informed by the others. What makes the goal challenging is that while the first two tasks involve bottom-up data-driven reasoning, recognition is inherently a high-level vision process.
We present a method that combines graphical models, pair-wise eigenvector-based clustering and Bayesian reasoning to learn an object-specific segmentation and tracking. Object recognition is performed by taking the maximum likelihood estimate (MLE) of these learned models. Preliminary results on the Personal Robotics robotic arm/segway videos show that the proposed approach is a significant improvement.

Authors: Dhruv Batra (CMU), Rahul Sukthankar, Tsuhan Chen (CMU)

Introducing Biodiversity in Enterprise Security [PDF]

Proteus explores the area of behavioral anomaly detection as it applies to end hosts. The current practice in enterprise networks is to use behavioral anomaly detectors at end hosts that are configured identically for all hosts, thus simplifying their management. However, such detectors typically rely on the notion of identifying statistical outliers. While one would expect most enterprise users to use the network in similar ways (in contrast to say home users), we show that behavior may greatly vary from user to user. Consequently, the cutoff threshold defining ``acceptable'' and ``unacceptable'' behavior should be defined for each user independently, a notion we call ``personalization hypothesis'' and which we extensively test in this project. Using end host data from 800 machines in two large corporate networks, we show that there is a great deal of diversity among users with respect to their outlier behavior and that this holds across a broad set of anomaly detectors. Biodiversity improves a host's ability to reduce missed detections, thus increasing the fraction of hosts within an enterprise that can detect anomalies. Interestingly, we find that the inherent variability in user outlier behavior also identifies the operating region where personalized detectors outperform homogeneous detectors the most.

Authors:

Toward an Optimal Social Network Defense against Sybil Attacks [PDF]

This poster presents SybilLimit, a new system to defend against Sybil attacks in collaborative, distributed systems. A Sybil attack is when a single malicious user creates multiple fake (Sybil) identities in order to "attack" the system. For example, a user with numerous fake identities can artificially build up his own reputation, ruin someone else's reputation, or grab more than his "fair share" of a shared system resource. SybilLimit extends the researchers' earlier SybilGuard work by dramatically reducing the number of Sybil identities accepted by the system. The SybilLimit defense is based on a novel use of real-world social networks.

Authors: Michael Kaminsky, Phil Gibbons, Haifeng Yu (National University of Singapore)

Applying Parallel Hardware to the Boolean Satisfiability Problem [PDF]

The Boolean Satisfiabilty Problem seeks to determine an assignment to the variables in a boolean function such that the function is satisfied (evaluates to true). This problem has many practical applications in the fields of logic synthesis, formal verification, model checking, and artificial intelligence. A small project in our lab has been investigating techniques for applying parallel hardware to the problem of boolean satisfiability. We will demonstrate an early prototype of a fine-grained algorithm running on a multicore system, as well as a coarse-grained technique running on multiple machines in a cluster.

Authors: Tamir Heyman (CMU), Limor Fix (IRP), Michael Kozuch (IRP), Edmund Clarke (CMU)

Optimal Replication for Multi-object Operations [PDF]

Masking failures is a key goal in distributed computing, and data replication is a well-known and widely-used technique to ensure data availability in the presence of failures. Previous research has focused on the availability of individual data objects (e.g., individual file blocks), ignoring the fact that a user-level task often needs to request multiple data objects (e.g., all the source files in a project). This poster shows both experimentally and analytically that the assignment of object replicas to machines has a subtle yet critical effect on the availability of such multi-object operations, even when the availability of individual objects remains the same. For example, popular assignments from the literature can have failure probabilities that differ by four orders of magnitude for the TPC-H decision support benchmark. We show a series of results regarding assignments that provide the best and the worst availability for user-level operations.

Authors: Phil Gibbons, Haifeng Yu (National University of Singapore)

Geographic Routing on Duty-Cycled Sensors [PDF]

Geographic routing is a useful and scalable point-to-point communication primitive for wireless sensor networks. However, previous work on geographic routing assumes that all the nodes in the network are awake during routing, overlooking the common deployment scenario where sensor nodes are duty-cycled to save energy. This poster presents our results on extending existing geographic routing algorithms to handle the highly dynamic networks resulting from duty-cycling. Moreover, we describe local sleep scheduling rules for sensor nodes that result in the emergence of a target routing latency for geographic routing on the nodes. Compared to previous approaches, this sleep scheduling algorithm significantly improves both routing latency and network lifetime.

Authors: Phil Gibbons, Suman Nath (Microsoft Research)

Self-organizing Wireless Access Networks [PDF]

The increased popularity of IEEE 802.11 WLANs has led to dense deployments in urban areas. Such high density leads to sub-optimal performance unless the interfering networks learn how to optimally share the spectrum. In this project we look into distributed algorithms that allow (i) multiple interfering 802.11 WLANs to select their operating frequency in a way that minimizes global interference, (ii) clients to choose their Access Point so that the bandwidth of all interfering networks is shared optimally, (iii) APs and clients to select their transmission power and Clear Channel Assessment (CCA) threshold so as to maximize the overall network capacity. The proposed algorithms optimize global network performance based on local information. They do not require explicit coordination among the wireless devices, and can thus operate in diverse environments with no single administrative authority.
This demo will showcase the effect of self-organization mechanisms on a 80-node Intel-based testbed deployment in University of Cambridge and a 30-node deployment at UC Riverside. Throughout the demo we will highlight the throughput benefits gained through self-organization, as well as their impact on the network topology for 802.11b/g and 802.11a.

Authors:

WLAN Conflict Graph Generation [PDF]

The rapid proliferation and dense deployment of wireless LANs has greatly increased RF interference, thus resulting in reduced system performance. To maximize performance, it is necessary to characterize the interference between radios without making overly simplistic assumptions about radio propagation. Moreover, it is important to recognize that the RF environment is highly dynamic. Keeping these constraints in mind, we propose a novel active probing approach to dynamically measure interference in enterprise WLAN environments. We use these and other measurements to annotate a conflict graph that (i) models interference of a radio transmitter on neighbouring nodes, and (ii) describes network performance using intuitive utility functions. Our approach captures interference between Access Points (APs), between clients and APs, and amongst clients. It does not require modifications to existing end-clients and consumes only a modest amount of resources. In our demo, we will use active probing to create a conflict graph in real time with little performance overhead and using legacy clients.

Authors: Nabeel Ahmed(University of Waterloo, Canada)

Self-organization for Wireless Mesh Networks [PDF]

Community wireless networks have been proposed to spread broadband network access to underpriviliged, underprovisioned and remote areas. Current research has only focused on operating the network on a single channel and finding better routing metrics for optimal traffic flow across the network. In this project, we argue that such an architecture does not optimally use the resources available, in that it does not utilize frequency diversity, interference knowledge in client's choices of gateways or perform load balancing.
We propose SWARM: a system that allows a community wireless network to optimally make use of the network resources. Finding an optimal network organization requires exhaustive enumeration and is highly complex. SWARM uses a set of four novel algorithms for node to gateway assignment, interference aware tree construction, performance modeling and frequency diversity that find network organizations that best utilize network resources while remaining practical. The evaluation of the SWARM system using simulations, as well as a deployed implementation in a 20 node network, demonstrate that SWARM can provide significant throughput gains by utilizing network resources effectively.

Authors:

Speculative Scheduling for Centralized WLANs [PDF]

In the past few years, there has been an increasing number of research proposals and industry attempts to maximize performance of 802.11-based enterprise wireless LANs through careful centralization. In this demo, we will showcase the gains one can achieve using centralized scheduling in some typical scenarios that are not handled efficiently by today's 802.11 medium access control mechanism (hidden terminal, exposed terminal). Further we will show that centralized schedulers are particularly well suited for dense deployment of thin APs, where the performance of the Distributed Coordination Function of IEEE 802.11 can degrade significantly.

Authors: Vivek Shrivastava(University of Wisconsin at Madison), ...

Make Your Uploads 5x Faster [PDF]

Last mile technology today, such as Cable and DSL, suffer from limited upstream capacity, thus restricting the type of services home users can engage in. In this project we have designed and implemented a file transfer protocol that exploits available wireless bandwidth in a neighbourhood to reach and make use of multiple wired uplinks, thus overcoming the upstream limitations of existing last mile technologies. The sender opportunistically broadcasts packets to neighbors, which intelligently use their aggregate upstream capacity to offer faster transfer completion times. The protocol itself is designed so as to minimize information redundancy while maximizing long-term throughput.
The demo compares a file upload using three schemes: 1) a single broadband connection; 2) unicast transmission to specific wireless neighbors; and 3) opportunistic wireless broadcast to all neighbors simultaneously.

Authors: Szymon Chachulski(MIT)

Distributed Detection and Containment of Outbreaks [PDF]

Distributed Detection and Inference (DDI) is an extremely sensitive, fully-distributed system for detecting weak network anomalies. End-hosts possessing "weak" local detectors collaborate by randomly gossiping information to other hosts where that information is used to update beliefs in a graphical probabilistic model to generate accurate conclusions about the network as a whole. Aside from being extremely sensitive, distributed inference is also robust to attack and failure of a central aggregation point, and possesses good scaling qualities without the presence of an explicit load-balancer. We will demonstrate DDI operating on a real 40 node network, and illustrate that it can detect and contain a difficult-to-detect stealthy worm attack.

Authors: Denver Dash, Eve Schooler, Jaideep Chandrashekar, Carl Livadas, John Mark Agosta, Tomas Singliar, Joohwan Kim, Geoff Weaver