CMU RI Seminar: Socially Guided Machine Learning

Andrea Thomaz
Post-Doctoral Associate
MIT

Abstract
There is a surge of interest in having robots leave the labs and factory floors to help solve critical issues facing our society, ranging from eldercare to education. A critical issue is that we will not be able to preprogram these robots with every skill they will need to play a useful role in society; robots will need the ability to interact and learn new things 'on the job' from everyday people. This talk introduces a paradigm, Socially Guided Machine Learning, that reframes the Machine Learning problem as a human-machine interaction, asking: How can systems be designed to take better advantage of learning from a human partner and the ways that everyday people approach the task of teaching?

In this talk I describe two novel social learning systems, on robotic and computer game platforms. Results from these systems show that designing agents to better fit human expectations of a social learning partner both improves the interaction for the human and significantly improves the way machines learn.

Sophie is a virtual robot that learns from human players in a video game via interactive Reinforcement Learning. A series of experiments with this platform uncovered and explored three principles of Social Machine Learning: guidance, transparency, and asymmetry. For example, everyday people were able to use an attention direction signal to significantly improve learning on many dimensions: a 50% decrease in actions needed to learn a task, and a 40% decrease in task failures during training.

On the Leonardo social robot, I describe my work enabling Leo to participate in social learning interactions with a human partner. Examples include learning new tasks in a tutelage paradigm, learning via guided exploration, and learning object appraisals through social referencing. An experiment with human subjects shows that Leo's social mechanisms significantly reduced teaching time by aiding in error detection and correction.

Lab Meeting 1 Feb 2007 (Any): Dynamic Maps for Long-Term Operation of Mobile Service Robots

Title: Dynamic Maps for Long-Term Operation of Mobile Service Robots
Authors: Peter Biber, Tom Duckett
Conference: Robotics: Science and System 2005 (RSS'05)
Local Copy: [PDF]

Abstract:

This paper introduces a dynamic map for mobile robots that adapts continuously over time. It resolves the stability-plasticity dilemma (the trade-off between adaptation to new patterns and preservation of old patterns) by representing the environment over multiple timescales simultaneously (5 in our experiments). A sample-based representation is proposed, where older memories fade at different rates depending on the timescale. Robust statistics are used to interpret the samples. It is shown that this approach can track both stationary and non-stationary elements of the environment, covering the full spectrum of variations from moving objects to structural changes. The method was evaluated in a five week experiment in a real dynamic environment. Experimental results show that the resulting map is stable, improves its quality over time and adapts to changes.

News: 'Sniffer-bot' algorithm helps robots seek scents

19:00 24 January 2007
NewScientist.com news service
Mason Inman

Moths are renowned for their ability to pick up a faint whiff of pheromones from faraway mates. Robots may soon match this feat with the help of a new mathematical method developed to help guide them toward a scent. In tests, the algorithm made virtual scent-hunter bots move just like moths do, snaking and spiralling toward their goal.

...

Massimo Vergassola at the Pasteur Institute in Paris, France, and colleagues created an algorithm that tells a robot how to move in order to gather as much olfactory information as possible. This allows it to home in on even the faintest of scents.

See the full article.
Journal reference: Nature (vol 445, p 406)

News: Web cam periscope

This little web cam accessory ($99) is like a periscope so you can look directly at the person you're video conferencing with... might be a fun re-make, links below to get you started - Link

Lab Meeting 25 Jan 2007(AShin):The Expectation Maximization Algorithm

Arthur:Frank Dellaert

College of Computing, Georgia Institute of Technology Technical Report

February 2002

Abstract:

This note represents my attempt at explaining the EM algorithm (Hartley, 1958;Dempster et al., 1977; McLachlan and Krishnan, 1997). This is just a slightvariation on Tom Minka’s tutorial (Minka, 1998), perhaps a little easier (or perhapsnot). It includes a graphical example to provide some intuition.

[Link]

Lab Meeting 25 Jan 2007 (ZhenYu):Psychophysiological control architecture for human-robot coordination-concepts and initial experiments

Nilanjan Sarkar
Dept. of Mech. Eng., Vanderbilt Univ., Nashville, TN;

Abstract:

The use of robots is expected to be pervasive in many spheres of society: in hospitals, homes, offices and battlefields, where the robots will need to interact and cooperate closely with a variety of people. The paper proposes an innovative approach to human-robot cooperation where the robot will be able to recognize the psychological state of the interacting human and modify its (i.e., robot's) own action to make the human feel comfortable in working with the robot. Wearable biofeedback sensors are used to measure a variety of physiological indices to infer the underlying psychological states (affective states) of the human. The eventual idea is to correlate the psychological states with the actions of the robot to determine which action(s) is responsible for a particular affective state. The robot controller will then modify that action if there is a need to alter the affective state. A concept of such a control architecture, a requirement analysis, and initial results from human experiments for stress detection are presented.

[link]

Lab Meeting 25 Jan 2007 (Casey): Contrast Context Histogram – A Discriminating Local Descriptor for Image Matching

From: The 18th International Conference on Pattern Recognition ( ICPR'06)
Title: Contrast Context Histogram – A Discriminating Local Descriptor for Image Matching
Author: Chun-Rong Huang, Chu-Song Chen and Pau-Choo Chung

Abstract:
This paper presents a new invariant local descriptor, contrast context histogram, for image matching. It represents the contrast distributions of a local region, and serves as a local distinctive descriptor of this region. Object recognition can be considered as matching salient corners with similar contrast context histograms on two or more images in our work. Our experimental results show that the developed descriptor is accurate and efficient for matching.

Paper Download: Link

CMU VASC talk: Subspectral Algorithms for Sparse Learning, Optimization & Inference

Baback Moghaddam
MERL
Monday, Jan 29, 3:30pm, NSH 1507

Subspectral Algorithms for Sparse Learning, Optimization & Inference

I will present a class of "subspectral" algorithms (i.e. sparse eigenvector techniques) for solving NP-hard combinatorial optimization problems in three general applied domains: (1) Supervised/unsupervised learning, in the traditional or orthodox sense (e.g. PCA & LDA), (2) Quadratic/Entropic Optimization (e.g. Least-Squares & MaxEnt) and (3) Inference, in the strict probabilistic/Bayesian sense (e.g. Automatic Relevance Determination and variational methods like Expectation Propagation). Subspectral algorithms for both exact (optimal) and greedy (approximate) solutions of these general sparse optimization problems are derived using analytic eigenvalue bounds. Specifically, an efficient "dual-pass" greedy algorithm is shown to yield near-optimal solutions for all possible cardinalities (at once) in a fraction of the time it takes for most continuous relaxation methods to find solutions of comparable quality for a single cardinality. I will present sample applications of subspectral optimization techniques in .sparse PCA. for feature selection (statistics), .sparse LDA. for classification (gene discovery), sparse kernel regression (robotics & control), sparse quadratic programming (portfolio optimization), graph model selection (sensor networks) as well as sparse Bayesian inference for computer vision (face recognition & OCR).

Bio:
Baback Moghaddam's research interests are in computational vision with a main focus on probabilistic visual learning. His related areas of interest and expertise include statistical modeling, Bayesian data analysis, machine learning and pattern recognition. He obtained his PhD in Electrical Engineering and Computer Science (EECS) from the Massachusetts Institute of Technology (MIT) in 1997 where he was a member of the Vision and Modeling Group at the MIT Media Laboratory where he developed a fully-automatic vision system which won DARPA's 1996 "FERET" Face Recognition Competition.

Dr. Moghaddam was the winner of the 2001 Pierre Devijver Prize from the International Association of Pattern Recognition for his "innovative approach to face recognition" and received the Pattern Recognition Society Award for "exceptional outstanding quality" for his journal paper "Bayesian Face Recognition." He currently serves on the editorial board of the journal Pattern Recognition and has contributed to numerous textbooks on image processing and computer vision including the core chapter in Springer-Verlag's latest biometric series, "Handbook of Face Recognition."

Dr. Moghaddam's past research included infrared (IR) image analysis for the Office of Naval Research (ONR), segmentation of synthetic aperture radar (SAR) imagery for MIT Lincoln Laboratory as well as designing a micro-gravity experiment for laser annealing of amorphous silicon which was flown aboard the US Space Shuttle in 1990.

http://www.merl.com/people/baback

CMU ML talk: Approximate inference using planar graph decomposition

Approximate inference using planar graph decomposition
by Amir Globerson and Tommi Jaakkola
NIPS 2006

A number of exact and approximate methods are available for inference calculations in graphical models. Many recent approximate methods for graphs with cycles are based on tractable algorithms for tree structured graphs. Here we base the approximation on a different tractable model, planar graphs with binary variables and pure interaction potentials (no external field). The partition function for such models can be calculated exactly using an algorithm introduced by Fisher and Kasteleyn in the 1960s. We show how such tractable planar models can be used in a decomposition to derive upper bounds on the partition function of non-planar models. The resulting algorithm also allows for the estimation of marginals. We compare our planar decomposition to the tree decomposition method of Wainwright et. al., showing that it results in a much tighter bound on the partition function, improved pairwise marginals, and comparable singleton marginals.

CMU ML talks: Greedy Layer-Wise Training of Deep Networks

Greedy Layer-Wise Training of Deep Networks
by Yoshua Bengio, Pascal Lamblin, Dan Popovici and Hugo Larochelle
NIPS 2006

Deep multi-layer neural networks have many levels of non-linearities allowing them to compactly represent highly non-linear and highly-varying functions. However, until recently it was not clear how to train such deep networks, since gradient- based optimization starting from random initialization appears to often get stuck in poor solutions. Hinton et al. recently introduced a greedy layer-wise unsupervised learning algorithm for Deep Belief Networks (DBN), a generative model with many layers of hidden causal variables. In the context of the above optimization problem, we study this algorithm empirically and explore variants to better understand its success and extend it to cases where the inputs are continuous or where the structure of the input distribution is not revealing enough about the variable to be predicted in a supervised task. Our experiments also confirm the hypothesis that the greedy layer-wise unsupervised training strategy mostly helps the optimization, by initializing weights in a region near a good local minimum, giving rise to internal distributed representations that are high-level abstractions of the input bringing better generalization.

News: Robot nurses ready for wards 'in three years'

http://www.thisislondon.co.uk/

Robot nurses could be bustling around hospital wards in as little as three years.

The mechanised "angels" being developed by EU-funded scientists, will perform basic tasks such as mopping up spillages, taking messages, and guiding visitors to hospital beds.

They could also distribute medicines and even monitor the temperature of patients remotely with laser thermometers.

...

He told The Engineer magazine: "The idea is not only to have mobile robots but also a full system of integrated information terminals and guide lights, so the hospital is full of interaction and intelligence.

"Operating as a completely decentralised network means that the robots can co-ordinate things between themselves, such as deciding which one would be best equipped to deal with a spillage or to transport medicine."

He said the robots could provide a valuable service guiding people around the hospital. A visitor would state the name of a patient at an information terminal and then follow a robot to the correct bedside.

...

See the full article.

MIT CSAIL PhD thesis: Robot Manipulation in Human Environments

Authors: Edsinger, Aaron
Advisor: Rodney Brooks
Issue Date: 16-Jan-2007

Abstract: Human environments present special challenges for robot manipulation. They are often dynamic, difficult to predict, and beyond the control of a robot engineer. Fortunately, many characteristics of these settings can be used to a robot's advantage. Human environments are typically populated by people, and a robot can rely on the guidance and assistance of a human collaborator. Everyday objects exhibit common, task-relevant features that reduce the cognitive load required for the object's use. Many tasks can be achieved through the detection and control of these sparse perceptual features. And finally, a robot is more than a passive observer of the world. It can use its body to reduce its perceptual uncertainty about the world. In this thesis we present advances in robot manipulation that address the unique challenges of human environments. We describe the design of a humanoid robot named Domo, develop methods that allow Domo to assist a person in everyday tasks, and discuss general strategies for building robots that work alongside people in their homes and workplaces. PDF

Paper: Experimental Characterization of Commercial Flash Ladar Devices

D. Anderson, H. Herman, and A. Kelly
International Conference of Sensing and Technology, November, 2005.

Abstract: Flash ladar is a new class of range imaging sensors. Unlike traditional ladar devices that scan a collimated laser beam over the scene, flash ladar illuminates the entire scene with diffuse laser light. Recently, several companies have begun offering demonstration flash ladar units commercially. In this work, we seek to characterize the performance of two such devices, examining the effects of target range, reflectance and angle of incidence, as well as mixed pixel effects.

PDF

CMU RI seminar: The neuroarchitecture of complex cognition

Marcel Adam Just
D. O. Hebb Professor of Psychology
Carnegie Mellon University

Recent findings in brain imaging, particularly fMRI, are beginning to reveal some of the fundamental properties of the organization of the cortical systems that underpin complex cognition. A set of operating principles govern the system organization, characterizing the system as a set of collaborating cortical centers that operate as a large-scale cortical network. Two of the network's critical features are that it is resource-constrained and dynamically-configured, with resource constraints and demands dynamically shaping the network topology. The operating principles are embodied in a cognitive neuroarchitecture, 4CAPS, consisting of a number of interacting computational centers that correspond to activating cortical areas. Each 4CAPS center is a hybrid production system, possessing both symbolic and connectionist attributes. 4CAPS models of several cognitive tasks (sentence comprehension, spatial problem solving, and complex multitasking) have been developed and compared to brain activation and behavioral results.

Congratulations to Casey Wang!

Congratulations! Casey Wang successfully defended his master thesis on "Hand Gesture Recognition using Adaboost with SIFT". Good job!

-Bob

IEEE ITSS Newsletter vol 8 nr 4, December 2006

In this issue you will find a lot of interesting material:
- ITSS related news; in particular many new initiatives from the ITS Society
- technical papers
- conference reports and announcements
- research overview focusing nomadic devices in the new intelligent vehicles

You can find the ITSS Newsletter at the IEEE ITSS official web site address:
http://www.ewh.ieee.org/tc/its/
or directly at:
http://www.its.washington.edu/itsc/v8n4.pdf

MIT CSAIL report: Latent-Dynamic Discriminative Models for Continuous Gesture Recognition

Authors: Morency, Louis-Philippe; Quattoni, Ariadna; Darrell, Trevor
Issue Date: 7-Jan-2007

Abstract: Many problems in vision involve the prediction of a class label for each frame in an unsegmented sequence. In this paper we develop a discriminative framework for simultaneous sequence segmentation and labeling which can capture both intrinsic and extrinsic class dynamics. Our approach incorporates hidden state variables which model the sub-structure of a class sequence and learn the dynamics between class labels. Each class label has a disjoint set of associated hidden states, which enables efficient training and inference in our model. We evaluated our method on the task of recognizing human gestures from unsegmented video streams and performed experiments on three different datasets of head and eye gestures. Our results demonstrate that our model for visual gesture recognition outperform models based on Support Vector Machines, Hidden Markov Models, and Conditional Random Fields.

PDF, PS

Lab meeting 12 Jan, 2007 (Stanley): A robocentric motion planner for dynamic environments using the velocity space

Author: E. Owen, L. Montano

From: IEEE/RSJ International Conference on Intelligent Robots and Systems. Oct. 9-15, Beijing, China.

Abstract:
This paper addresses a method to optimize the robot motion planning in dynamic environments, avoiding the moving and static obstacles while the robot drives towards the goal. The method maps the dynamic environment into a model in the velocity space, computing the times to potential collision and potential escape and the associated robot velocities. The problem of finding a trajectory to the goal is stated as a constrained nonlinear optimization problem. The initial seed trajectory for the optimization is directly generated in the velocity space using the model built. The method is applied to robots which are subject to both kinematic constraints (i.e. involving the configuration parameters of the robot and their derivatives), and dynamic constraints, (i.e. the constraints imposed by the acceleration/deceleration capabilities). Some experimental results are discussed.

Link

News: iRobot Introduces the iRobot Create!

A robust, programmable Robot platform that invites you to stretch your imagination

iRobot, a pioneering robot company that has sold millions of iRobot Roomba vacuuming robots, has introduced a remarkable robot platform that fills a major gap.

The iRobot Create is a dependable, rugged and versatile robot base that can be used for uncounted robotics hobby and research applications. It includes a selection of software routines developed for iRobot’s commercial appliance bots and a well-engineered, robust chassis designed for longevity. Many will consider this to be a DIY-roboticist’s dream come true.

Here, we offer a summary of the iRobot Create’s features in this First Look, initial comments by contributing editor Dan Lynch (who has been playing with one for a few days as of this post) and an interesting overview of applications designed for this new robot platform by iRobot employees worldwide. Stay tuned for an in-depth feature article in the Summer 2007 issue of Robot.

The iRobot Create comes fully assembled. It has 32 built-in sensors, two powered wheels, a castor (and optional 4th castor wheel), 10 pre-programmed behaviors, an expandable input/ouput port for custom sensors and actuators, a cargo bay with mounting points and a tailgate for ballast. This new bot platform works with optional accessories such as the iRobot Command Module, iRobot Roomba Virtual Wall units, the self-charging home base, and iRobot Roomba standard remote. You can use the Roomba rechargeable battery options or standard alkaline batteries. You’ll need a computer with a serial port (USB connectivity is expected soon) and Microsoft Windows XP, Linux or Mac OS X. - Link

Related: iRobot "Create" - Educational robot - Link

News: The First Kyosho Athlete Humanoid Cup!

The launch of the MANOI AT01 humanoid robot kit, featured in our Winter 2006 issue, was an instant success with many stores in Japan selling out of stock within a few days of its September introduction. So it was no surprise when a large number of customers turned out with their robots fully assembled and customized to compete in the first Kyosho Athletics Humanoid Cup event on December 10th in the trendy Omotesando Hills complex in Tokyo.

The inaugural Kyosho Athlete Humanoid Cup event was of keen interest to robot fans everywhere. The images below tell the story of how this exciting competition unfolded.

Left to right: Dr. GIY's MANOI AT01, Sugiura's AT01 and a silver MANOI PF01 were exhibited at a MANOI launch press conference in September 2006.

The competitions, watched by standing-room only crowds, included 5-meter sprints against the clock with both R/C and autonomous divisions, plus two minute demonstrations/performances scored by an expert panel of judges.

See the full article.

paper: Predictive Mover Detection and Tracking in Cluttered Environments

L. Navarro, C. Mertz, and M. Hebert
Proc. of the 25th. Army Science Conference, November, 2006.
PDF

Abstract:This paper describes the design and experimental evaluation of a system that enables a vehicle to detect and track moving objects in real-time. The approach investigated in this work detects objects in LADAR scan lines and tracks these objects (people or vehicles) over time. The system can fuse data from multiple scanners for 360° coverage. The resulting tracks are then used to predict the most likely future trajectories of the detected objects. The predictions are intended to be used by a planner for dynamic object avoidance. The perceptual capabilities of our system form the basis for safe and robust navigation in robotic vehicles, necessary to safeguard soldiers and civilians operating in the vicinity of the robot.

paper: Bootstrap learning of foundational representations

B.J. Kuipers, P. Beeson, J. Modayil, J. Provost
Connection Science, 2006 - Taylor & Francis
PDF

Abstract: To be autonomous, intelligent robots must learn the foundations of commonsense knowledge from their own sensorimotor experience in the world. We describe four recent research results that contribute to a theory of how a robot learning agent can bootstrap from the “blooming buzzing confusion” of the pixel level to a higher-level ontology including distinctive states, places, objects, and actions. This is not a single learning problem, but a lattice of related learning tasks, each providing prerequisites for tasks to come later. Starting with completely uninterpreted sense and motor vectors, as well as an unknown environment, we show how a learning agent can separate the sense vector into modalities, learn the structure of individual modalities, learn natural primitives for the motor system, identify reliable relations between primitive actions and created sensory features, and can define useful control laws for homing and path-following. Building on this framework, we show how an agent can use to self-organizing maps to identify useful sensory featurs in the environment, and can learn effective hill-climbing control laws to define distinctive states in terms of thos features, and trajectoryfollowing control laws to move from one distinctive state to another. Moving on to place recognition, we show how an agent can combine unsupervised learning, map-learning, and supervised learning to achieve high-performance recognition of places from rich sensory input. And finally, we take the first steps toward learning an ontology of objects, showing tha a bootstrap learning robot can learn to individuate objects through motion, separating them from the static environment and from each other, and learning properties that will be useful for classification. These are four key steps in a much larger research enterprise that lays the foundation for human and robot commonsense knowledge.

News: Wow Wee Unveils Robopanda

Cute, but not cuddly robot will be the company's most charming and interactive part of its 2007 line.

One of the highlights of AI, Steve Spielberg's melancholy look at robots in the distant future, was a little autonomous robot bear, "Teddy," that served as friend and companion to the film's main character David the "mech" boy robot. The bear could walk, talk interact and show real intelligence and affection. Now Wow Wee is apparently ripping a page from that movie's script to introduce the new Robopanda.

Part of Wow Wee's 2007 line of robot toys, the $229 Robopanda is swathed in plastic instead of cuddly fur, and is roughly 19.5 inches tall (standing), 11 inches wide and 6 inches deep. Wow Wee officials, however, promise a level of robot/human interaction scarcely seen in previous robot toys.

Designed for children (and, perhaps, adults) ages 4 and up, the 8-pound Robopanda will be covered with eight touch sensors and, using infrared and stereo sensors, should be able to avoid obstacles, track objects and locate sound sources. It will also tell stories and recognize and interact with its own companion: a plush-panda toy that will ship with the robot.

The robo-mammal will also feature, Wow Wee execs said, "incredible movement," using its nine "ultra-quiet" motors (and one tilt sensor) to sit, crawl, walk (on all fours), roll over and hug. Unlike previous Wow Wee robots, Robopanda is not expected to ship with a remote. Its artificial intelligence may also be greater than previous Wow Wee products: "Robpanda responds with mood-specific behaviors…based on [user] interaction," Wow Wee notes in a recent press release.

See the full article.

News: Spyke Wi-Fi Spy Robot Debuts at CES 2007

Consumer robots could also be a big topic at the CES 2007. Here is a newcomer. French MECCANO introduces the Spyke robot in the United States under the ERECTOR brand.

The Spyke robot is controlled via a PC over Wi-Fi. Spyke has a webcam and moves on a rubber band. Professional robots can climb stairs with such a drive mechanism.

See the full article.

News: Gates says day of the home-help robot is near

James Randerson, science correspondent
Friday January 5, 2007
The Guardian

An office worker checks her home-gadget webpage from her work computer. The tasks she set for her home robots in the morning have all been completed: washing and ironing, vacuuming the lounge and mowing the lawn.

She orders dinner from the kitchen chefbot - sushi today, using a recipe from a Japanese website - then checks her elderly mother's house. The companionbot has given mum her medicine and helped her out of bed and into a chair.

This is the vision of the future offered by Bill Gates who, in the latest issue of Scientific American, argues that the robotics industry is on the cusp of a big expansion. He likens the current state of robotic technology to the situation in the fledgling computer industry when he and his fellow entrepreneur Paul Allen launched Microsoft in the mid-1970s.

See the full article.

Lab meeting 5 Jan, 2007 (Nelson): Motion–Egomotion Discrimination and Motion Segmentation from Image-Pair Streams

David Demirdjian and Radu Horaud
[LINK]

Computer Vision and Image Understanding
Volume 78 , Issue 1 (April 2000)
Special issue on robust statistical techniques in image understanding

Pages: 53 - 68

Abstract:

Given a sequence of image pairs we describe a method that segments the observed scene into static and moving objects while it rejects badly matched points.We show that, using a moving stereo rig, the detection of motion can be solved in a projective framework and therefore requires no camera calibration. Moreover the method allows for articulated objects. First we establish the projective framework enabling us to characterize rigid motion in projective space. This characterization is used in conjunction with a robust estimation technique to determine egomotion. Second we describe a method based on data classification which further considers the non-static scene points and groups them into several moving objects. Third we introduce a stereo-tracking algorithm that provides the point-to-point correspondences needed by the algorithms. Finally we show some experiments involving a moving stereo head observing both static and moving objects.

Lab meeting 5 Jan, 2007(Atwood): Plan-view trajectory estimation with dense stereo background models

Title: Plan-view trajectory estimation with dense stereo background models
Auther: Darrell, T.; Demirdjian, D.; Checka, N.; Felzenszwalb, P.

from ICCV 2001.

Abstract:

In a known environment, objects may be tracked in multiple views using a set of background models. Stereo-based models can be illumination-invariant, but often have undefined values which inevitably lead to foreground classification errors. We derive dense stereo models for object tracking using long-term, extended dynamic-range imagery, and by detecting and interpolating uniform but unoccluded planar regions. Foreground points are detected quickly in new images using pruned disparity search. We adopt a “late-segmentation” strategy, using an integrated plan-view density representation. Foreground points are segmented into object regions only when a trajectory is finally estimated, using a dynamic programming-based method. Object entry and exit are optimally determined and are not restricted to special spatial zones


Link

News: Robot, My Slippers Please

Home sensors, long-distance health monitors and other gadgets help seniors remain independent.

May 2006

RI-MAN isn’t your average caregiver. The pale-green, 220-pound robot is a mass of wiring, metal and computer chips. It was created in Japan as an eventual high-tech alternative to costly home-health services and nursing-home care.

Although you can’t order your own RI-MAN or other home-care robot yet, you can buy many other assistive-technology devices that enable older adults with various ailments to continue to live in their own homes. Such devices include home sensors that monitor a person’s day-to-day activities and special goggles that help the visually impaired to see. These products are part of tech companies’ response to the new demographics: a rising number of seniors, families scattered around the globe and grown children with full-time careers who care for elderly parents. Here are some examples of what’s available now.

See the full article.

Lab meeting 28 Dec, 2006 (Jim): Unified Inverse Depth Parametrization for Monocular SLAM

Unified Inverse Depth Parametrization for Monocular SLAM
Montiel etal., RSS 2006

PDF
A.J.Davison's website

Abstract:
      Recent work has shown that the probabilistic SLAM approach of explicit uncertainty propagation can succeed in permitting repeatable 3D real-time localization and mapping even in the ‘pure vision’ domain of a single agile camera with no extra sensing. An issue which has caused difficulty in monocular SLAM however is the initialization of features, since information from multiple images acquired during motion must be combined to achieve accurate depth estimates. This has led algorithms to deviate from the desirable Gaussian uncertainty representation of the EKF and related probabilistic filters during special initialization steps.
      In this paper we present a new unified parametrization for point features within monocular SLAM which permits efficient and accurate representation of uncertainty during undelayed initialisation and beyond, all within the standard EKF (Extended Kalman Filter). The key concept is direct parametrization of inverse depth, where there is a high degree of linearity. Importantly, our parametrization can cope with features which are so far from the camera that they present little parallax during motion, maintaining sufficient representative uncertainty that these points retain the opportunity to ‘come in’ from infinity if the camera makes larger movements. We demonstrate the parametrization using real image sequences of large-scale indoor and outdoor scenes.

Lab meeting 28 Dec, 2006 (Any): Sonar Sensor Interpretation

Title: Sonar Interpretation Learned from Laser Data
Authors: S. Enderle, G. Kraetzschmar, S. Sablatnog and G. Palm
From: 1999 Third European Workshop on Advanced Mobile Robots, 1999. (Eurobot '99)
Links: [Paper 1][Paper 2][Paper 3]
Abstract:
Sensor interpretation in mobile robots often involves an inverse sensor model, which generates hypotheses on specific aspects of the robot's environment based on current sensor data. Building inverse sensor models for sonar sensor assemblies is a particularly difficult problem that has received much attention in past years. A common solution is to train neural networks using supervised learning. However; large amounts of training data are typically needed, consisting e.g. of scans of recorded sonar data which are labeled with manually constructed teacher maps. Obtaining these training data is an error-prone and time-consuming process. We suggest that it can be avoided, if an additional sensor like a laser scanner is also available which can act as the feeding signal. We have successfully trained inverse sensor models for sonar interpretation using laser scan data. In this paper; we describe the procedure we used and the results we obtained.

Lab meeting 28 Dec, 2006 (Leo): Square Root SAM

Square Root SAM: Simultaneous Localization and Mapping via Square Root Information Smoothing

Frank Dellaert

Robotics: Science and Systems, 2005

Abstract— Solving the SLAM problem is one way to enable
a robot to explore, map, and navigate in a previously unknown
environment. We investigate smoothing approaches as a viable
alternative to extended Kalman filter-based solutions to the
problem. In particular, we look at approaches that factorize either
the associated information matrix or the measurement matrix
into square root form. Such techniques have several significant
advantages over the EKF: they are faster yet exact, they can be
used in either batch or incremental mode, are better equipped
to deal with non-linear process and measurement models, and
yield the entire robot trajectory, at lower cost. In addition,
in an indirect but dramatic way, column ordering heuristics
automatically exploit the locality inherent in the geographic
nature of the SLAM problem.
In this paper we present the theory underlying these methods,
an interpretation of factorization in terms of the graphical model
associated with the SLAM problem, and simulation results that
underscore the potential of these methods for use in practice.


[Link]

Lab meeting 15 Dec, 2006 (Casey): Estimating 3D Hand Pose from a Cluttered Image

Title: Estimating 3D Hand Pose from a Cluttered Image
Authors: Vassilis Athitsos and Stan Scalaroff
(CVPR 2003)

Abstract:
A method is proposed that can generate a ranked list of
plausible three-dimensional hand configurations that best
match an input image. Hand pose estimation is formulated
as an image database indexing problem, where the closest
matches for an input hand image are retrieved from a large
database of synthetic hand images. In contrast to previous
approaches, the system can function in the presence of
clutter, thanks to two novel clutter-tolerant indexing methods.
First, a computationally efficient approximation of
the image-to-model chamfer distance is obtained by embedding
binary edge images into a high-dimensional Euclidean
space. Second, a general-purpose, probabilistic line matching
method identifies those line segment correspondences
between model and input images that are the least likely to
have occurred by chance. The performance of this cluttertolerant
approach is demonstrated in quantitative experiments
with hundreds of real hand images.

Paper download: [Link]

Lab meeting 15 Dec, 2006 (YuChun): Modeling Affect in Socially Interactive Robots

Author:
Rachel Gockley, Reid Simmons, and Jodi Forlizzi

Proc. of the 15th IEEE International Symposium on Robot and Human Interactive Communication (RO-MAN06), September, 2006.

Abstract:
Humans use expressions of emotion in a very social manner, to convey messages such as “I'm happy to see you” or “I want to be comforted,” and people's long-term relationships depend heavily on shared emotional experiences. We believe that for robots to interact naturally with humans in social situations they should also be able to express emotions in both short-term and long-term relationships. To this end, we have developed an affective model for social robots. This generative model attempts to create natural, human-like affect and includes distinctions between immediate emotional responses, the overall mood of the robot, and long-term attitudes toward each visitor to the robot. This paper presents the general affect model as well as particular details of our implementation of the model on one robot, the Roboceptionist.


[Link]

[Thesis Oral] A Market-Based Framework for Tightly-Coupled Planned Coordination in Multirobot Teams

Author:
Nidhi Kalra
Robotics Institute
Carnegie Mellon University

Abstract:
This thesis explores the coordination challenges posed by real-world multirobot domains that require planned tight coordination between teammates throughout execution. These domains involve solving a multi-agent planning problem in which the actions of robots are tightly coupled. Because of uncertainty in the environment and the team, they also require persistent tight coordination between teammates throughout execution.

This thesis proposes an approach to these problems in which the complexity and strength of the coordination adapt to the difficulty of the problem. Our approach, called Hoplites, is a market-based framework that selectively injects pockets of complex coordination into a primarily distributed system by enabling robots to purchasing each other's participation in tightly-coupled plans over the market. We discuss how it is widely applicable to real-world problems because it is general, computationally feasible, scalable, operates under uncertainty, and improves solutions with new information. Experiments show that our approach significantly outperforms existing coordination methods.

Lab meeting 8 Dec, 2006 (Chihao): Particle filtering algorithms for tracking an acoustic source in a reverberant environment

Author:
Ward, D.B. Lehmann, E.A. Williamson, R.C.
Dept. of Electr. & Electron. Eng., Imperial Coll. London, UK

From: Speech and Audio Processing, IEEE Transactions

Abstract:

Traditional acoustic source localization algorithms attempt to find the current location of the acoustic source using data collected at an array of sensors at the current time only. In the presence of strong multipath, these traditional algorithms often erroneously locate a multipath reflection rather than the true source location. A recently proposed approach that appears promising in overcoming this drawback of traditional algorithms, is a state-space approach using particle filtering. In this paper we formulate a general framework for tracking an acoustic source using particle filters. We discuss four specific algorithms that fit within this framework, and demonstrate their performance using both simulated reverberant data and data recorded in a moderately reverberant office room (with a measured reverberation time of 0.39 s). The results indicate that the proposed family of algorithms are able to accurately track a moving source in a moderately reverberant room.

[Link]

Lab meeting 8 Dec, 2006 (AShin): Learning and Inferring Transportation Routines

Author:L. Liao, D. Fox, and H. Kautz

Proc. of the National Conference on Artificial Intelligence (AAAI-04)
Outstanding Paper Award

Abstract
This paper introduces a hierarchical Markov model that can learn and infer a user's daily movements through the community. The model uses multiple levels of abstraction in order to bridge the gap between raw GPS sensor measurements and high level information such as a user's mode of transportation or her goal. We apply Rao-Blackwellised particle filters for efficient inference both at the low level and at the higher levels of the hierarchy. Significant locations such as goals or locations where the user frequently changes mode of transportation are learned from GPS data logs without requiring any manual labeling. We show how to detect abnormal behaviors (\eg\ taking a wrong bus) by concurrently tracking his activities with a trained and a prior model. Experiments show that our model is able to accurately predict the goals of a person and to recognize situations in which the user performs unknown activities.

[Link]

No Polit, No Problem?

[origional link]

The promise is fantastic: new generations of remote-controlled aircraft could soon be flying in civilian airspace, performing all sorts of useful tasks.The reality is that a lack of radio frequencies to control the planes and serious concerns over their safety are going to keep them grounded for years to come.
Surprisingly, given the commercial hopes it has for civil unmanned aerial vehicles (UAVs), the aviation industry has failed to obtain the radio frequencies it needs to control them - and it will be 2011 before it can even begin to lobby for space on the radio spectrum. What's more, none of the world's aviation authorities will allow civil UAVs to fly in their airspace without a reliable system for avoiding other aircraft - and the industry has not yet even begun developing such a system. Experts say this could take up to seven years.
Dedicated frequencies are handed out at the International Telecommunications Union's World Radiocommunications Conference.But no one in the UAV industry had applied for any new frequencies.If UAVs are to mingle safely with other civilian aircraft, the industry needs to develop a safe, standardised collision avoidance system. This is complicated because aviation regulators demand that if UAVs are to have access to civil airspace, they must be "equivalent" in every way to regular planes.The problem for now is that aviation regulators have yet to define precisely what they mean by "equivalent", so UAV makers are not yet willing to commit themselves to developing collision-avoidance technology."A crewless aircraft on a collision course must behave as if it had a pilot on board"
On the brighter side, last week the UN's International Civil Aviation Organization said its navigation experts would meet in early 2007 to consider regulations for UAVs in civil airspace.
however, it will be meaningless unless the industry can obtain the necessary frequencies to control the planes and feed images and other sensor data back to base, says Bowker. "The lack of robust, secure radio spectrum is a show-stopper."

Context Aware Computing, Understanding and Responding to Human Intention

[Original Link]

Ted Selker

Abstract

This talk will demonstrate that Artificial intelligence can competently Improve human interaction with systems and even each other in a myriad of natural scenarios. Humans work to understand and react to each others intentions. The context aware computing group at the MIT Media lab has demonstrated that across most aspects of our life, computers can do this too. The groups demonstrations range from car to office kitchen to and even bed. The goal is to show that human intentions can be recognized considered and responded to appropriately by computer systems. Understanding and acting appropriately to intentions requires more than good sensors, it requires understanding of the value of the input. The context aware demonstrations therefore rely completely on models of what the system can do, what the tasks are that can be performed and what is known about the user . These models of system task and user form a central basis for deciding when and how to respond in a specific situation.

Dr. Ted Selker is an Associate Professor at the MIT Media, the Director of the Context Aware Computing Lab, the MIT director of The Voting Technology Project and the Counter Intelligence/ Design Intelligence special interest group on domestic and product-design of the future. Ted's work strives to demonstrate that peoples intentions can be recognized and respected by the things we design. Context aware computing creates a world in which peoples desires and intentions cause computers to help them. This group is recognized for its creating environments that use sensors and artificial intelligence to create so-called "virtual sensors"; adaptive models of users to create keyboard less computer scenarios. Ted's Design Intelligence work has used technology rich platforms such as kitchens to examine intention based design., Ted's work is also applied to developing and testing user experience technology and security architectures for recording and voter intentions securely and accurately.

Prior to joining MIT faculty in November 1999, Ted was an IBM fellow and directed the User Systems Ergonomics Research lab. He has served as a consulting professor at Stanford University, taught at Hampshire, University of Massachusetts at Amherst and Brown Universities and worked at Xerox PARC and Atari Research Labs.

Ted's research has contributed to products ranging from notebook computers to operating systems. His work takes the form of prototype concept products supported by cognitive science research. He is known for the design of the TrackPoint in-keyboard pointing device found in many notebook computers, and many other innovations at IBM. Ted's technologies are often featured in national and international news media.

Ted is work has resulted in award winning products, numerous patents, papers and is often featured by the press. And was co recipient of computer science policy leader awarded for Scientific American 50 in 2004 and the American Association For People with Disabilities Thomas Paine Award for his work on voting technology.

Lab meeting 1 Dec, 2006 (ZhenYu):Detecting Social Interaction of Elderly in a Nursing Home Environment

Author:
Datong Chen, Jie Yang, Robert Malkin, and Howard D. Wactlar

Abstract:
Social interaction plays an important role in our daily lives. It is one of the most important indicators of physical or mental changes in aging patients. In this paper, we investigate the problem of detecting social interaction patterns of patients in a skilled nursing facility. Our studies consist of both a “wizard of Oz” study and an experimental study of various sensors and detection models for detecting and summarizing social interactions among aging patients and caregivers. We first simulate plausible sensors using human labeling on top of audio and visual data collected from a skilled nursing facility. The most useful sensors and robust detection models are determined using the simulated sensors. We then present the implementation of some real sensors based on video and audio analysis techniques and evaluate the performance of these implementations in detecting interaction. We conclude the paper with discussions and future work.

Download:[link]

Lab meeting 1 Dec, 2006 (Nelson): Randam Sample Consensus: A Paradigm for Model Fitting with Applications to Image Analysis and Automated Cartography

Martin A. Fischler and Robert C. Bolles
SRI International

Communication of the ACM
June 1981 Volume 24 Number 6

LINK

Abstrct:

A new paradigm, Random Sample Consensus
(RANSAC), for fitting a model to experimental data is
introduced. RANSAC is capable of interpreting/
smoothing data containing a significant percentage of
gross errors, and is thus ideally suited for applications
in automated image analysis where interpretation is
based on the data provided by error-prone feature
detectors. A major portion of this paper describes the
application of RANSAC to the Location Determination
Problem (LDP): Given an image depicting a set of
landmarks with known locations, determine that point
in space from which the image was obtained. In
response to a RANSAC requirement, new results are
derived on the minimum number of landmarks needed
to obtain a solution, and algorithms are presented for
computing these minimum-landmark solutions in closed
form. These results provide the basis for an automatic
system that can solve the LDP under difficult viewing

News: Scientists Try to Make Robots More Human

November 22, 2006

George the robot is playing hide-and-seek with scientist Alan Schultz. What's so impressive about robots playing children's games? For a robot to actually find a place to hide, and then hunt for its human playmate is a new level of human interaction. The machine must take cues from people and behave accordingly.

This is the beginning of a real robot revolution: giving robots some humanity.

"Robots in the human environment, to me that's the final frontier," said Cynthia Breazeal, robotic life group director at the Massachusetts Institute of Technology. "The human environment is as complex as it gets; it pushes the envelope."

Robotics is moving from software and gears operating remotely - Mars, the bottom of the ocean or assembly lines - to finally working with, beside and even on people.

"Robots have to understand people as people," Breazeal said. "Right now, the average robot understands people like a chair: It's something to go around."

See the full article.

News: Pleo, the sensitive robot

Behold the majesty of Pleo. It's a robot, coming out in the second quarter of 2007, that exhibits emotional reactions to its surroundings. So far, companion robots have been a big flop in the market, but Pleo maker Ugobe hopes to succeed by pricing it for around $250, lower than other companion robots, and giving consumers ways to program it.

Don't be fooled by the eyes; the vision system is in its nose.

When it's in a good mood, the Pleo wags it tail back and forth. It also makes a sort of mooing sound, which is appropriate given that the robot is modeled after the Camarasaurus, a cow-like dinosaur that roamed the Americas in the Jurassic period. A paleontologist helped Ugobe come up with the design.

There's a lot going on under the Pleo's rubbery skin. The robot contains six microprocessors and more than 150 gears. It also has a memory card slot where the sun don't shine.

Credit: Michael Kanellos/CNET News.com

See the related link & video.

News: Robot Senses Damage, Learns to Walk Again

November 17, 2006

—It may look like a metallic starfish, but scientists say this robot might have more in common with a newborn human.

The four-legged machine is a prototype "resilient robot" with the ability to detect damage to itself and alter its walking style in response.

Josh Bongard, an assistant professor of computer science at the University of Vermont in Burlington, and his colleagues created the robot as part of a NASA pilot project working on technology for the next generation of planetary rovers.

While people and animals can easily compensate for injuries, even a small amount of damage can ground NASA machinery entirely.

See the full article.

CMU ML talk: Machine Learning and Human Learning

Speaker: Prof. Tom Mitchell, CMU
http://www.cs.cmu.edu/~tom
Date: November 20
Time: 12:00 noon

For schedules, links to papers et al, please see the web page:
http://www.cs.cmu.edu/~learning/

Abstract:

For the past 30 years, researchers studying machine learning and researchers studying human learning have proceeded pretty much independently. We now know enough about both fields that it is time to re-ask the question: "How can studies of human learning and studies of machine learning inform one another?" This talk will address this question by briefly covering some of the key facts we now understand about both machine learning and human learning, then examining in some detail several specific types of machine learning which may provide surprisingly helpful models for understanding aspects of human learning (e.g., reinforcement learning, cotraining).

Lab meeting 17 Nov, 2006 (Jim): Real-Time Simultaneous Localisation and Mapping with a Single Camera

Real-Time Simultaneous Localisation and Mapping with a Single Camera
Andrew J. Davison, ICCV 2003

PDF, homepage

Abstract:
      Ego-motion estimation for an agile single camera moving through general, unknown scenes becomes a much more challenging problem when real-time performance is required rather than under the off-line processing conditions under which most successful structure from motion work
has been achieved. This task of estimating camera motion from measurements of a continuously expanding set of self-mapped visual features is one of a class of problems known as Simultaneous Localisation and Mapping (SLAM) in the robotics community, and we argue that such real-time mapping research, despite rarely being camera-based, is more relevant here than off-line structure from motion methods due to the more fundamental emphasis placed on propagation of uncertainty.
      We present a top-down Bayesian framework for single-camera localisation via mapping of a sparse set of natural features using motion modelling and an information-guided active measurement strategy, in particular addressing the difficult issue of real-time feature initialisation via a factored sampling approach. Real-time handling of uncertainty permits robust localisation via the creating and active measurement of a sparse map of landmarks such that regions can be re-visited after periods of neglect and localisation can continue through periods when few features are visible. Results are presented of real-time localisation for a hand-waved camera with very sparse prior scene knowledge and all processing carried out on a desktop PC.

Jaron Lanier forecasts the future

* 16 November 2006
* NewScientist.com news service
* Jaron Lanier

In the next 50 years, computer science needs to achieve a new unification between the inside of the computer and the outside. The inside is still governed by the mid-20th century approach – that is, every program must have defined inputs and outputs in order to function.

The outside, however, encounters the real world and must analyse data statistically. Robots must assess a terrain in order navigate it. Language translation programs must make guesses in order to function. Because the interface to the outside world involved approximation, it is also capable of adjusting itself to improve the quality of approximation. But the inside of a computer must adhere to protocols to function at all, and therefore cannot evolve automatically.

See the full article.

Eric Horvitz forecasts the future

* 8 November 2006
* NewScientist.com news service
* Eric Horvitz

Computation is the fire in our modern-day caves. By 2056, the computational revolution will be recognised as a transformation as significant as the industrial revolution. The evolution and widespread diffusion of computation and its analytical fruits will have major impacts on socioeconomics, science and culture.

Within 50 years, lives will be significantly enhanced by automated reasoning systems that people will perceive as "intelligent". Although many of these systems will be deployed behind the scenes, others will be in the foreground, serving in an elegant, often collaborative manner to help people do their jobs, to learn and teach, to reflect and remember, to plan and decide, and to create. Translation and interpretation systems will catalyse unprecedented understanding and cooperation between people. At death, people will often leave behind rich computational artefacts that include memories, reflections and life histories, accessible for all time.

Robotic scientists will serve as companions in discovery by formulating theories and pursuing their confirmation.

See the full article.

Rodney Brooks forecasts the future

* 18 November 2006
* NewScientist.com news service
* Rodney Brooks

Show a two-year-old child a key, a shoe, a cup, a book or any of hundreds of other objects, and they can reliably name its class - even when they have never before seen something that looks exactly like that particular key, shoe, cup or book. Our computers and robots still cannot do this task with any reliability. We have been working on this problem for a while. Forty years ago the Artificial Intelligence Laboratory at the Massachusetts Institute of Technology appointed an undergraduate to solve it over the summer. He failed, and I failed on the same problem in my 1981 PhD.

In the next 50 years we can solve the generic object recognition problem. We are no longer limited by lack of computer power, but we are limited by a natural risk aversion to a problem on which many people have foundered in the past few decades.

See the full article.

Lab meeitng 17 Nov., 2006 (Any): World modeling for an autonomous mobile robot using heterogenous sensor information

Title: World modeling for an autonomous mobile robot using heterogenous sensor information

Local Copy: [Here]
Related Link: [Here]

Author: Klaus-Werner Jorg
From: Robotics and Autonomous Systems, 1995

Abstract:
An Autonomous Mobile Robot (AMR) has to show both goal-oriented behavior and reflexive behavior in order to be considered fully autonomous. In a classical, hierarchical control architecture these behaviors are realized by using several abstraction levels while their individual informational needs are satisfied by associated world models. The focus of this paper is to describe an approach which utilizes heterogenous information provided by a laser-radar and a set of sonar sensors in order to achieve reliable and complete world models for both real-time collision avoidance and local path planning. The approach was tested using MOBOT-IV, which serves as a test-platform within the scope of a research project on autonomous mobile robots for indoor applications. Thus, the experimental results presented here are based on real data.

Lab meeitng 17 Nov., 2006 (Stanley): Policies Based on Trajectory Libraries

Author:
Martin Stolle

Abstract:
We present a control approach that uses a library of trajectories to establish a global control law or policy. This is an alternative to methods for finding global policies based on value functions using dynamic programming and also to using plans based on a single desired trajectory. Our method has the advantage of providing reasonable policies much faster than dynamic programming can provide an initial policy. It also has the advantage of providing more robust and global policies than following a single desired trajectory. Trajectory libraries can be created for robots with many more degrees of freedom than what dynamic programming can be applied to as well as for robots with dynamic model discontinuities. Results are shown for the “Labyrinth” marble maze, both in simulation as well as a real world version. The marble maze is a difficult task which requires both fast control as well as planning ahead.

Link:
ICRA 2006
Thesis proposal

News: Google and Microsoft aim to give you a 3-D world

By Brad Stone
Newsweek

Nov. 20, 2006 issue - The sky over San Francisco is cerulean blue as you begin your descent into the city from 2,000 feet. As you pass over the southern hills, the skyline of the Financial District rises into view. On the descent into downtown, familiar skyscrapers form an urban canyon around you; you can even see the trolley tracks running down the valley formed by Market Street. But then a little pop-up box next to the Bay Bridge explains that an accident has just occurred on the west-ern span, and a thick red line indicates the resulting traffic jam along the highway. A banner ad for Emeryville, Calif., firm ZipRealty hangs incongruously in the air over the Transamerica Pyramid. You are actually staring at your PC screen, not out an airplane window.

Virtual Earth 3D, the online service unveiled last week by Microsoft, is both incomplete (only 15 cities are depicted in 3-D) and imperfect (some of the buildings are shrouded in shadow, and you need a powerful PC running Windows XP or the new Vista to use it). But it is also the start of something potentially big: the 3-D Web. Traditional Web pages give us text, photos and video, unattached to real-world context. Now interactive mapping programs like Google Earth let us zoom around the globe on our PCs and peer down at the topography captured by satellites and aerial photographers. Both Google Earth and Microsoft's Virtual Earth are hugely popular and have been downloaded more than 100 million times each.

See the full article.

[Folks, we can build 4D maps of the real world in a much faster and more reliable way, right? -Bob]

Lab Meeting 10 Nov talk (Casey): Object Class Recognition Using Multiple Layer Boosting with Heterogeneous Features

Title: Object Class Recognition Using Multiple Layer Boosting with Heterogeneous Features
Authors: Wei Zhang, Bing Yu, Gregory J. Zelinsky, Dimitris Samaras
(CVPR 2005)

Abstract:
We combine local texture features (PCA-SIFT), global features (shape context), and spatial features within a single
multi-layer AdaBoost model of object class recognition.
The first layer selects PCA-SIFT and shape context features and combines the two feature types to form a strong classifier.
Although previous approaches have used either feature type to train an AdaBoost model, our approach is the first to combine these complementary sources of information into a single feature pool and to use Adaboost to select
those features most important for class recognition.
The second layer adds to these local and global descriptions information about the spatial relationships between features.
Through comparisons to the training sample, we first find the most prominent local features in Layer 1, then capture
the spatial relationships between these features in Layer 2.
Rather than discarding this spatial information, we therefore use it to improve the strength of our classifier. We compared our method to [4, 12, 13] and in all cases our approach outperformed these previous methods using a popular
benchmark for object class recognition [4].
ROC equal error rates approached 99%. We also tested our method using a dataset of images that better equates the complexity
between object and non-object images, and again found that our approach outperforms previous methods.

Download: [link]

Robot learns to grasp everyday chores- BRIAN D. LEE

Original link

From left, graduate students Ashutosh Saxena and Morgan Quigley and Assistant Professor Andrew Ng were part of a large effort to develop a robot to see an unfamiliar object and ascertain the best spot to grasp it.

Stanford scientists plan to make a robot capable of performing everyday tasks, such as unloading the dishwasher. By programming the robot with "intelligent" software that enables it to pick up objects it has never seen before, the scientists are one step closer to creating a real life Rosie, the robot maid from The Jetsons cartoon show.

"Within a decade we hope to develop the technology that will make it useful to put a robot in every home and office," said Andrew Ng, an assistant professor of computer science who is leading the wireless Stanford Artificial Intelligence Robot (STAIR) project.

"Imagine you are having a dinner party at home and having your robot come in and tidy up your living room, finding the cups that your guests left behind your couch, picking up and putting away your trash and loading the dishwasher," Ng said.

Cleaning up a living room after a party is just one of four challenges the project has set out to have a robot tackle. The other three include fetching a person or object from an office upon verbal request, showing guests around a dynamic environment and assembling an IKEA bookshelf using multiple tools.

Developing a single robot that can solve all these problems takes a small army of about 30 students and 10 computer science professors—Gary Bradski, Dan Jurafsky, Oussama Khatib, Daphne Koller, Jean-Claude Latombe, Chris Manning, Ng, Nils Nilsson, Kenneth Salisbury and Sebastian Thrun.

From Shakey to Stanley and beyond

Stanford has a history of leading the field of artificial intelligence. In 1966, scientists at the Stanford Research Institute built Shakey, the first robot to combine problem solving, movement and perception. Flakey, a robot that could wander independently, followed. In 2005, Stanford engineers won the Defense Advanced Research Projects Agency (DARPA) Grand Challenge with Stanley, a robot Volkswagen that autonomously drove 132 miles through a desert course.

The ultimate aim for artificial intelligence is to build a robot that can create and execute plans to achieve a goal. "The last serious attempt to do something like this was in 1966 with the Shakey project led by Nils Nilsson," Ng said. "This is a project in Shakey's tradition, done with 2006 technology instead of 1966 AI technology."

To succeed, the scientists will need to unite fragmented research areas of artificial intelligence including speech processing, navigation, manipulation, planning, reasoning, machine learning and vision. "There are these disparate AI technologies and we'll bring them all together in one project," Ng said.

The true problem remains in making a robot independent. Industrial robots can follow precise scripts to the point of balancing a spinning top on a blade, he said, but the problem comes when a robot is requested to perform a new task. "Balancing a spinning top on the edge of a sword is a solved problem, but picking up an unfamiliar cup is an unsolved problem," Ng explained.

His team recently designed an algorithm that allowed STAIR to recognize familiar features in different objects and select the right grasp to pick them up. The robot was trained in a computer-generated environment to pick up five items—a cup, pencil, brick, book and martini glass. The algorithm locates the best place for the robot to grasp an object, such as a cup's handle or a pencil's midpoint. "The robot takes a few pictures, reasons about the 3-D shape of the object, based upon computing the location, and reaches out and grasps the object," Ng said.

In tests with real objects, the robotic arm picked up items similar to those with which it trained, such as cups and books, as well as unfamiliar objects including keys, screwdrivers and rolls of duct tape. To grasp a roll of duct tape, the robot employs an algorithm that evaluates the image against all prior strategies. "The roll of duct tape looks a little like a cup handle and also a little bit like a book," Ng said. The program formulates the best location to clutch based on a combination of all the robot's prior experiences and tells the arm where to go. "It would be a hybrid, or a combination of all the different grasping strategies that it has learned before," Ng said.

The word "robot" originates from a Slavic word meaning "toil," and robots may soon reduce the amount of drudgery in our daily lives. "I think if we can have a robot intelligent enough to do these things, that will free up vast amounts of human time and enable us to go to higher goals," Ng said.

Funding for the project has come from the National Science Foundation, DARPA and industrial technology companies Intel, Honda, Ricoh and Google.

Brian D. Lee is a science writing intern with the Stanford News Service.

[FRC seminar] Learning Robot Control Policies Using the Critique of Teacher

Speaker:
Brenna Argall
Ph.D. Candidate
Robotics Institute

Abstract:
Motion control policies are a necessary component of task execution on mobile robots. Their development, however, is often a tedious and exacting procedure for a human programmer. An alternative is to teach by demonstration; to have the robot extract its policy from the example executions of a teacher. With such an approach, however, most of the learning burden is typically placed with the robot. In this talk we present an algorithm in which the teacher augments the robot's learning with a performance critique, thus shouldering some of the learning burden. The teacher interacts with the system in two phases: first by providing demonstrations for training, and second by offering a critique on learner performance. We present an application of this algorithm in simulation, along with preliminary implementation on a real robot system. Our results show improved performance with teacher critiquing, where performance is measured by both execution success and efficiency.

Speaker Bio:
Brenna is currently a third year Ph.D. candidate in the Robotics Institute at Carnegie Mellon University, affiliated with the CORAL Research Group. Her research interests lie with robot autonomy and heterogeneous team coordination, and how machine learning may be used to build control policies which accomplish these tasks. Prior to joining the Robotics Institute, Brenna investigated functional MRI brain imaging in the Laboratory of Brain and Cognition at the National Institutes of Health. She received her B.S. in Mathematics from Carnegie Mellon in 2002, along with minors in Music and Biology.

NEWS: Robot guide makes appearance at Fukushima hospital

A robot with the ability to recognize speech and show hospital visitors the way to consultation rooms and wards has made its debut at Aizu Central Hospital in Aizuwakamatsu.

When the robot is asked to pinpoint a location, it projects a three-dimensional image showing the route to the destination through a projector on its head, and prints out a map from a built-in printer, which it hands to the visitor.

AIZUWAKAMATSU, Fukushima
November 5, 2006

See the full article

[CMU VASC Seminar]Real Time 3D Surface Imaging and Tracking for Radiation Therapy

VASC Seminar Series

Speaker: Maud Poissonnier, Vision RT

Time: Thursday, 11/9

Title: Real Time 3D Surface Imaging and Tracking for Radiation Therapy

Abstract:

Radiation Therapy involves the precise delivery of high energy X-rays to
tumour tissue in order to treat cancer. The current challenge is to ensure
that the radiation is delivered to the correct target location, thus
reducing the volume of normal tissue irradiated and potentially enabling
the escalation of dose. This may be achieved through the exploitation of a
combination of imaging technologies.

Vision RT has developed a 3D surface imaging technology which can image
the entire 3D surface of a patient quickly and accurately. This relies on
close range stereo photogrammetric techniques using pairs of stereo
cameras. Registration algorithms are employed to match surface data
acquired from the same patient in different positions. High speed tracking
techniques have also been developed to allow tracking of regions at speeds
of approximately 20 fps.

AlignRT(r) is Vision RT's patient setup and surveillance system which is
now in use at a variety of clinics. The system acquires 3D surface data
during simulation or imports reference contours from diagnostic Computed
Tomography (CT). It then images the patient prior to treatment and
computes any 3D movement required to correct the patient's position. The
system is also able to monitor any patient movement during treatment. We
will finish by presenting work-in-progress systems which allow real time
tracking of breathing motion to facilitate 4D CT reconstruction and
respiratory gated radiotherapy.

Bio:

Dr. Maud Poissonnier was educated in France until she decided to
explore Great Britain in 1995. She first graduated at Heriot-Watt
University in Edinburgh with an MSc in Reservoir Evaluation and Management
(Petroleum Engineering). She then joined the Medical Vision Lab (Robotics
Research Group) at the University of Oxford. She obtained her EPSRC-funded
DPhil in 2003 under the supervision of Sir Prof. Mike Brady in the area of
x-ray mammography image processing using physics-based modelling. After a
post-doctoral research on a multimodality, Grid enabled platform for
tele-mammography, she moved very slightly East (to London) and joined
Vision RT Ltd in April 2005 where she is presently Senior Software
Engineer.

[NEWS]Walking Partner Robot helps old ladies cross the street

Nomura Unison Co., Ltd. has developed a walking-assistance robot that perceives and responds to its environment. The machine, called Walking Partner Robot, was developed with the cooperation of researchers from Tohoku University. It will be unveiled to the general public at the 2006 Suwa Area Industrial Messe on October 19 in Suwa, Nagano prefecture.
The robot is equipped with a system of sensors that detect the presence of obstacles, stairs, etc. while monitoring the motion and behavior of the user. Three sensors monitor the status of the user while detecting and measuring the distance to potential obstacles, and two angle sensors measure the slope of the path in front of the machine. The robot responds to these measurements with voice warnings and by automatically applying brakes when necessary.
Walking Partner Robot is essentially a high-tech walker designed to support users as they walk upright, preventing them from falling over. The user grasps a set of handles while pushing the unmotorized 4-wheeled robot, which measures 110 (H) x 70 (W) x 80 (D) cm and weighs 70 kilograms (154 lbs).
Walking Partner Robot is the second creation from the team responsible for the Partner Ballroom Dance Robot, which includes Tohoku University robotics researchers Kazuhiro Kosuge and Yasuhisa Hirata. The goal was to apply the Partner Ballroom Dance Robot technology, which perceives the intended movement and force of human footsteps, to a robot that can play a role in the realm of daily life. The result is a machine that can perceive its surroundings and provide walking assistance to the elderly and physically disabled.
The developers, who also see potential medical rehabilitation applications, aim to develop indoor and outdoor models of the robot. The company hopes to make the robot commercially available soon at a price of less than 500,000 yen (US$4,200) per unit.

A Braille Writing Tutor to Combat Illiteracy in Developing Communities

Title: A Braille Writing Tutor to Combat Illiteracy in Developing Communities
Speaker: Nidhi Kalra, Tom Lauwers
Date/Time/Location: Tuesday, November 7th, 2006, 11am, NSH 3305
Abstract:
We present the Braille Writing Tutor project, an initiative sponsored by TechBridgeWorld to combat the high rates of illiteracy among the blind in developing communities using an intelligent tutoring system. Developed in collaboration withe Mathru Educational Trust for the Blind in Bangalore, India, the tutor uses a novel input device to capture students' activity on a slate and stylus and uses a range of scaffolding techniques and Artificial Intelligence to teach writing skills to both beginner and advanced students. We conducted our first field study from August to September 2006 at the Mathru School to evaluate its feasibility and impact in a real educational setting. The tutor was met with great enthusiasm by both the teachers and the students and has already demonstrated a concrete impact on the students' writing abilities. Our study also highlights a number of important areas for future research and development which we invite the community to explore and investigate with us.

For more information and videos:
http://www.cs.cmu.edu/~nidhi/brailletutor.html

Speaker Bio:
Nidhi Kalra is a fifth year Ph.D. student at the Robotics Institute. She is keenly interested in applying technology to sustainable development and in understanding related public policy issues. Nidhi hopes to start a career in this field after completing her Ph.D. Her thesis area of research is in multirobot coordination and she is advised by Dr. Anthony Stentz. She has an MS in Robotics from the RI and received her BS in computer science from Cornell University in 2002. Nidhi is a native of India.

Tom Lauwers is a fourth year Ph.D. student at the Robotics Institute. He has a long-standing interest in educational robotics, as both a participant in programs like FIRST and later as a designer of a robotics course and education technology. He is currently studying curriculum development and evaluation and hopes that his study of the educational sciences will help him design better and more useful educational technologies. Tom received a BS in Electrical Engineering and a BS in Public Policy from CMU.

[The ML Lunch talk ]Boosting Structured Prediction for Imitation Learning

Speaker: Nathan Ratliff, CMU
http://www.cs.cmu.edu/~ndr

Title: Boosting Structured Prediction for Imitation Learning

Venue: NSH 1507

Date: November 06

Time: 12:00 noon


Abstract:

The Maximum Margin Planning (MMP) algorithm solves imitation learning
problems by learning linear mappings from features to cost functions in
a planning domain. The learned policy is the result of minimum-cost
planning using these cost functions. These mappings are chosen so that
example policies (or trajectories) given by a teacher appear to be lower
cost (with a loss-scaled margin) than any other policy for a given
planning domain. We provide a novel approach, MMPBoost, based on the
functional gradient descent view of boosting that extends MMP by
``boosting'' in new features. This approach uses simple binary
classification or regression to improve performance of MMP imitation
learning, and naturally extends to the class of structured maximum
margin prediction problems. Our technique is applied to navigation and
planning problems for outdoor mobile robots and robotic legged
locomotion.

In this talk, I will first provide an overview of the MMP approach to
imitation learning, followed by an introduction to our boosting
technique for learning nonlinear cost functions within this framework. I
will finish with a number of experimental results and a sketch of how
stuctured boosting algorithms of this sort can be derived.

[FRC Seminar] Learning-enhanced Market-based Task Allocation for Disaster Response

Speaker:
E. Gil Jones
Ph.D. Candidate
Robotics Institute

Abstract:
This talk will introduce a learning-enhanced market-based task allocation system for disaster response domains. I model the disaster response domain as a team of robots cooperating to extinguish a series of fires that arise due to a disaster. Each fire is associated with a time-decreasing reward for successful mitigation, with the value of the initial reward corresponding to task importance, and the speed of decay of the reward determining the urgency of the task. Deadlines are also associated with each fire, and penalties are assessed if fires are not extinguished by their deadlines. The team of robots aims to maximize summed reward over all emergency tasks, resulting in the lowest overall damage from the series of fires.

In this talk I will first describe my implementation of a baseline market-based approach to task allocation for disaster response. In the baseline approach the allocation respects fire importance and urgency, but agents do a poor job of anticipating future emergencies and are assessed a high number of penalties. I will then describe two regression-based approaches to learning-enhanced task allocation. The first approach, task-based learning, seeks to improve agents' valuations for individual tasks. The second method, schedule-based learning, tries to quantify the
tradeoff between performing a given task or not performing the task and retaining the flexibility to better perform future tasks. Finally, I will compare the performance of the two learning methods and the baseline approach over a variety of parameterizations of the disaster response domain.

Speaker Bio:
Gil is a second year Ph.D. student at the Robotics Institute, and is co-advised by Bernardine Dias and Tony Stentz. His primary interest is market-based multi-robot coordination. He received his BA in Computer Science from Swarthmore College in 2001, and spent two years as a software engineer at Bluefin Robotics - manufacturer of autonomous underwater vehicles - in Cambridge, Mass.

Lab meeitng 1 Nov., 2006 (Vincent): Fitting a Single Active Appearance Model Simultaneously to Multiple Images

In this talk, I will present the following 2 papers.

Title :
Fitting a Single Active Appearance Model Simultaneously to Multiple Images

Author :
Changbo Hu, Jing Xiao, Iain Matthews, Simon Baker, Jeff Cohn and Takeo Kanade
The Robotics Institute,CMU

Origin :
BMVC 2004

Abstract :
Active Appearance Models (AAMs) are a well studied 2D deformable model. One recently proposed extension of AAMs to multiple images is the Coupled-View AAM. Coupled-View AAMs model the 2D shape and appearance of a face in two or more views simultaneously. The major limitation of Coupled-View AAMs, however, is that they are specific to a particular set of cameras, both in geometry and the photometric responses. In this paper, we describe how a single AAM can be fit to multiple images, captured simultaneously by cameras with arbitrary geometry and response functions. Our algorithm retains the major benefits of Coupled-View AAMs:~the integration of information from multiple images into a single model, and improved fitting robustness.



Title :
Real-Time Combined 2D+3D Active Appearance Models

Author :
Jing Xiao, Simon Baker, Iain Matthews, and Takeo Kanade
The Robotics Institute,CMU

Origin :
CVPR 2004

Abstract :
Active Appearance Models (AAMs) are generative models commonly used to model faces. Another closely related type of face models are 3D Morphable Models (3DMMs). Although AAMs are 2D, they can still be used to model 3D phenomena such as faces moving across pose. We first study the representational power of AAMs and show that they can model anything a 3DMM can, but possibly require more shape parameters. We quantify the number of additional parameters required and show that 2D AAMs can generate model instances that are not possible with the equivalent 3DMM. We proceed to describe how a non-rigid structure-from-motion algorithm can be used to construct the corresponding 3D shape modes of a 2D AAM. We then show how the 3D modes can be used to constrain the AAM so that it can only generate model instances that can also be generated with the 3D modes. Finally, we propose a real-time algorithm for fitting the AAM while enforcing the constraints, creating what we call a "Combined 2D+3D AAM."

[NEWS]iRobot Unveils New Technology for Simultaneous Control of Multiple Robots

iRobot Corp. today released the first public photo of a new project in development, code named Sentinel. This innovative new networked technology will allow a single operator to simultaneously control and coordinate multiple semi-autonomous robots via a touch-screen computer. Funded by the U.S. Army's Small Business Innovation and Research (SBIR) program, the Sentinel technology includes intelligent navigation capabilities that enable the robots to reach a preset destination independently, overcoming obstacles and other challenges along the way without intervention from an operator.Sentinel’s capability will allow warfighters and first responders to use teams of iRobot® PackBot® robots to conduct surveillance and mapping, therefore rendering dangerous areas safe without ever setting foot in a hostile environment.

[Full]

Lab meeitng 1 Nov., 2006 (Chihao): Microphone Array Speaker Localizers Using Spatial-Temporal Information

author:
Sharon Gannot and Tsvi Gregory Dvorkind

from:
EURASIP Journal on Applied Signal Processing
Volume 2006

abstract:

A dual-step approach for speaker localization based on a microphone array is addressed in this paper. In the first stage, which is not the main concern of this paper, the time difference between arrivals of the speech signal at each pair of microphones is estimated. These readings are combined in the second stage to obtain the source location. In this paper, we focus on the second stage of the localization task. In this contribution, we propose to exploit the speaker’s smooth trajectory for improving the current position estimate. Three localization schemes, which use the temporal information, are presented. The first is a recursive form of the Gauss method. The other two are extensions of the Kalman filter to the nonlinear problem at hand, namely, the extended Kalman filter and the unscented Kalman filter. These methods are compared with other algorithms, which do not make use of the temporal information. An extensive experimental study demonstrates the advantage of using the spatial-temporalmethods. To gain some insight on the obtainable performance of the localization algorithm, an approximate analytical evaluation, verified by an experimental study, is conducted. This study shows that in common TDOA-based localization scenarios—where the microphone array has small interelement spread relative to the source position—the elevation and azimuth angles can be accurately estimated, whereas the Cartesian coordinates as well as the range are poorly estimated.

Link