April, 2003
We remain focused on our objectives to develop complete, effective and scalable software for autonomous robot teams. Demonstrate robot teams with integrated perception, reasoning, learning, communication and cooperative strategies that solve complex multi-agent tasks.
The following web sites are relevant to this program:
| http://www.cs.cmu.edu/~coral/MARS | The main site for details on our MARS project and copies of our presentations and reports. |
| http://www.cs.cmu.edu/~coral | Our main website. This site contains links to our publications, movies, and specific project web sites. |
| http://www.cs.cmu.edu/~robosoccer | Our main robot soccer website. This site contains links to our team pages, which in turn contain specific information on each domain we have investigated in robot soccer. |
| http://www.cs.cmu.edu/~robosoccer/segway | Our new CM-RMP Segway RMP robot soccer web site. |
| http://www.americanopen03.org | The First RoboCup American Open'03 website for the competition we hosted in May 2003. |
| http://www.cs.cmu.edu/~coral/publications | Our main publications website. This site contains links to all of our publications from the CORAL research group. |
The people working under the MARS program consists of faculty, graduate students, and undergraduate students as follows:
| Faculty: | Prof. Manuela Veloso, Dr. Brett Browning |
| Graduates: | Michael Bowling, James Bruce, Paul Carpenter, Scott Lenser, Patrick Riley, Douglas Vail |
| Undergraduates: | Michael Sokolsky, Juan Fasola, Sonia Chernova, Jennifer Lin, Ling Xu, Paul Harada, Betsy Ricker, David Rozner |
Adversarial multi-robot problems, where teams of robots compete with one another, require the development of approaches that span all levels of control and integrate algorithms ranging from low-level robot motion control, through to planning, opponent modeling, and multi-agent learning. Our research is primarily focused on addressing these multi-robot control issues in task-driven adversarial environments. In particular, we demonstrate much of our work within the context of robot soccer; a task-driven domain involving unknown adversaries in a dynamic and uncertain environment.
Under the MARS program, we continue to focus on the key challenges for building individually skilled autonomous robots and successful multi-robot teams to operate in uncertain, adversarial domains. These challenges range from single robot issues of localization, real-time visual sensing, robust tracking, high-speed navigation control, and automatic segmentation and recognition of environmental regimes through to multi-robot team issues of adaptive team strategy, robust cooperation through communication in the face of uncertainty, and team learning. All of these issues are important to the successful operation of multi-robot teams.
During the last quarter, our focus has been on developing and extending our techniques and algorithms for responding to an adversarial team and testing them at robot soccer events against unknown opponents. In particular, we have recently competed at the International RoboCup 2003, held in Italy during July, and at the First RoboCup American Open, which we hosted in May at Carnegie Mellon University. We proved very successful at the American Open, where we won in both the small-size and the Sony AIBO leagues. Likewise at RoboCup we finished in the 4th place position in both leagues. We describe our research advancements leading up to these events in greater detail below.
A second component of our recent work has focused on the new Segway RMP platform. We have now received our new robot base, and successfully teleoperated the robot using our existing small-size robot soccer software. A video of this work is available here. We describe our plans for this platform in more detail in the future plan section.
We hosted the American Open at Carnegie Mellon University from April 30 to May 4. The American Open was a regional event open to teams from the American continent, ie. USA, Canada, Mexico, Brazil, Chile etc. As such it proved very successful with:
| League | Nr. Teams | Champions | Results Link |
| Sony AIBO | 8 | CMPack'03 | Results |
| Small-size | 10 | CMDragons'03 | Results |
| Simulation | 15 (some remote) | UvATrilearn (U. of Amsterdam) | Results |
| Simulation Coach | 6 | IUST | Results |
| RoboCup Junior and U-League | 10 (+3 U-league) | Taipei 1, Danville Ironbots, Canada A.I, | Results |
| Rescue Robot Demonstration | 5 | NA | NA |
A vast majority of these teams were competiting in a robot soccer competition for the first time. All of the robot teams were from the American continent (ie. US, Canada, Mexico, Brazil, Chile etc), while the soccer simulation league was open to all competitors via remote participation. In conjunction with the robot soccer competition, we held a series of workshops and panel discussions after each day of competition to help focus on the research contributions from the event.
In conjunction with hosting the American Open, we participated in the event with three teams as preperationfor RoboCup 2003. The teams were our Sony AIBO team CM-Pack'03, our small-size team CMDragons'03, and our Simulator Coach team OWL'03. Both our Sony AIBO team and our small-size team won their respective leagues. Most interestingly, a number of teams within the Sony AIBO league and small-size leagues had made use of our 2002 source code releases. Indeed, our finalist opponents in the Sony AIBO team from Georgia Intitute of Technology, developed their software from our 2002 source code release. In the small-size league the team from Chile used robots and software developed from out small-size 2002 release.
In our previous work under the MARS program we have developed a fixed team-plan coordination system for controlling a team of robots operating in dynamic, adversarial environments. We call each fixed-team plan a 'play', due to its similarity to its sporting counterpart. Team behavior then consists of a reportoire of plays collected into a playbook. Each play consists of a sequence of parameterized behaviors assigned to each role on the team, where roles are then assigned dynamically at run-time to robots to execute them. Each play is written as a text file using a specialized play language. One of the key parts of our approach is the use of adaptation in the play selection process. Our recent work has focused on extending and validating our play adaptation system. We have now modified our weight updating mechanisms to use a sleeping experts approach. We have also improved the credit assignment mechanisms to more accurately credit the plays that lead to goals being scored. We are currently focused on evaluating the performance of our play-based approach through analyzing the log-files recorded during our games against a variety of unknown opponents at RoboCup 2003. Published results are forthcoming.
We have previously developed a vision-based obstacle avoidance system based on the use of color segmentation using our CMVision library to identify obstacles from the ground surface. The vision algorithm uses our Sony AIBOs and color vision library CMVision ( http://www.cs.cmu.edu/~jbruce/cmvision) developed earlier within the MARS program.
We have now extended our previous work to greatly enhance the useability of our approach in practise. In addition to efficiency improvements, which mean our avoidacne algorithms can be used on-line even with the limited processing capabilities of the Sony AIBO, we have provided multiple spatial representations to support different uses of the local map built by the robot as it navigates through the world. In addition to improvements to the radial spatial representation we developed previously, we have also included a cartesian 'occupancy grid' based representation. The former is useful for navigation techniques like vector field histogramming, the latter for A*-like planning techniques. Thus, the obstacle avoidance system can be queried in a manner suitable to the navigation techniques used.
We have previously developed a novel technique called Sensor Resetting Localization, for handling poorly modelled movement updates in Monte-Carlo Localization methods. Examples of such poorly modelled movement updates include 'teleportation', where a robot is physically moved by an external agent in an unpredictable way. In our recent work, we have further extending our technique for sensor-resetting using a combination of Metropolis sampling with an energy-minimization approach to provide better incorporation of sensor data in a more rigorous manner. This work is still relatively new, and will be investigated further.
Although many researchers have investigated ways to build robot behaviors, few have done so within an adversarial environment. Adversarial tasks are unique in that the state space for the problem explodes with the presence of other agents (since from a decision-theoretic viewpoint, actions should be conditioned on the full state of the world). From a behavioral perspective, this means that each robot must have a much more extensive behavior reportoire than it might otherwise have needed in world in which it alone is the only animate entitiy. We have developed a coherent state-machine based approach to behavior generation. Although many components of our approach are not novel, their combination is. Indeed, our behavior architectures now run on all of our robot platforms, which have quite different capabilities. Moreover, we have learnt many lessons on how behaviors should be implemented so as to avoid oscillation, perceptual aliasing issues and looping, while maintaining good performance. We are in the process of compiling these lessons for future publication.
Building on our previous multi-robot learning work under the MARS program, we have extending our multi-agent learning approach, called WoLF for Win or Learn Fast, to include continuous state spaces to arrive at a new algorithm we call Gra-WoLF. Gra-WoLF, short for Gradient-WoLF, utilizes linear function approximation, parameterized stochastic policy representations with policy gradient hill climbing (hence Gradient-WoLF), with the WoLF algorithm. We have applied the Gra-WoLF algorithm to a adversarial multi-robot task, called KeepOut, using our small-size robots (CMDragons) built under the MARS program. KeepOut involves many of the aspects of multi-robot systems that make learning a challenging task including non-negligible latencies, noise, and perceptual aliasing. We have validated our approaches both using a realistic simulator and using our real robot platform. We have recently published this work in the form of an IJCAI conference paper.
We have received our Segway RMP platform and have additionally bought a Segway HT platform. During the last quarter we have transitioned some of our small-size software to allow us to teleoperate the robot via a laptop and wavelan to as a first step towards developing infrastructure for the platform. It also provides us with a means for gaining experience with the unique dynamics of the vehicle. Our footage of some of the experiments performed with the robot are shown on our Segway web site at http://www.cs.cmu.edu/~robosoccer/segway.
For the next quarter, leading up to the PI meeting in late September, we will be focused on building the software and hardware infrastructure to enable the Segway RMP to play a soccer-like game outdoors on a grass field. In terms of hardware, we will choose the appropriate cameras, computers, and communications, and will also build a special purpose kicking mechanism to enable the Segway to kick the ball effectively. In terms of software, we will begin by transitioning our vision techniques developed on our other platforms. Our goals will be to develop basic ball following skills initially to provide the basis for transitioning our behaviors and soccer skills developed on the AIBO and Small-size platforms.
Our primary goal for the Sony AIBO platform will focus on RoboCup 2003, to be held in Padua, Italy. Following this, during the fall period, we will be concentrating on extending our recent research accomplishments with the Sony AIBOs, and the products of the work now going on for RoboCup 2003. We will continue our investigations into localization using our newly developed metropolis based sampling technique. We will continue to develop our vision-based obstacle avoidance techniques in conjunction with our work on the Segway RMP. We will also work on extending our methods for shared-world modelling for robots with high communications latencies, and local vision. We will also be teaching a new class in the fall period based on our work with the Sony AIBO's. All course materials, and relevant software, will be made available on-line. This includes a repackaged version of our robot soccer software to lower the startup costs for other researchers who wish to use the Sony AIBO platform.
Our primary goal for the small-size paltform will be focused on RoboCup 2003, to be held in Padua, Italy. In addition to further developing the team capabilities in now ongoing work, we will be carefully logging each game to gain valuable data to analyze the performance of our adaptable play-based strategy engine when exposed to a range of unknown opponents of different capabilities. These data logs will also provide a good basis for furthering the development of our opponent modelling techniques, particularly as data of many different teams is traditionally difficult to come by. We will also make our data logs available to the community via the web. During the fall period, we will also focus on extending our RRT based fast planning techniques, including examining direct performance comparisons to more traditional techniques such as A*. We will also focus on making avaiable our simulation engine which has now been used for an extended period during our development.
All of our software and research publications will continue to be available on the internet. A list of our recent publications is shown below. A number of our programs are already on-line and have been used by a number of researchers world-wide. Indeed, at the recent American Open, and RoboCup, many teams had made use of our software releases to help improve their capabilities. Additionally, our fast vision library CMVision, is now being used within the Stage/Player framework. As part of this process, we will be transfering our work to the Segway RMP platform, which we will then release to the growing Segway RMP community for use and perusal.
All of our work is available on-line at http://www.cs.cmu.edu/~coral/download/. Our publications are avilable at http://www.cs.cmu.edu/~coral/publications/, with the most recent, relevant publications listed below.
Our software includes:
We expect to be release a new series of software during the fall period following RoboCup 2003.