CMRoboBits is a course offered at Carnegie Mellon University that introduces students to all the concepts needed to create a complete intelligent robot. In particular, the course focuses on the areas of perception, cognition, and action by using the Sony AIBO robot as the focus for the programming assignments. This course shows how an AIBO and its software resources make it possible for students to investigate and work with an unusually broad variety of AI topics within a single semester. While material presented in this article describes using AIBOs as the primary platform, the concepts presented in the course are not unique to the AIBO and can be applied on different kinds of robotic hardware.

TheCMPACK ’04 team follows our past teams CMPack’03, CMPack’02, CMPack’01, CMPack’00, CMTrio’99, and CMTrio’98 [7, 5, 10, 9]. In our research this year, we explored the capabilities of the new ERS-7 robotic platform; developed new ways of modeling the world, such as using accelerometer data to detect when the robot is entangled; and maintained our focus on robust behaviors and cooperation. In this paper, we will focus in detail on several recent components of the team.

Robots performing tasks constantly encounter changing environmental conditions. These changes in the environment vary from the dramatic, such as rearrangement of furniture, to the subtle, such as a burnt out light bulb or a different carpeting. We do not recognize many of these changes, especially subtle changes, but robots do. These changes often lead to the failure of robots. In this thesis, we develop an algorithm for detecting these changes. Traditional sensor models do not capture all of the dependencies in the sensor data and are not capable of detecting all types of signal changes while maintaining a strong probabilistic foundation. This thesis corrects these shortcomings. We show how detecting the current conditions in which the robot is operating can lead to increased performance and lower failure rates. The methods in this thesis are tested on real tasks performed by a real robot, namely a Sony AIBO robot.

We extend our previous work on a classification algorithm for time series. Given time series produced by different underlying generating processes, the algorithm predicts future time series values based on past time series values for each generator. Unlike many algorithms, this algorithm predicts a distribution over future values. This prediction forms the basis for labelling part of a time series with the underlying generator that created it given some labelled exam piles. The algorithm is robust to a wide variety of possible types of changes in signals including mean shifts, amplitude changes, noise changes, period changes, and changes in signal shape. We improve upon the speed of our previous approach and show the utility of the algorithm for discriminating between different states of the robot/environment from robotic sensor signals.

CM-Pack’01 came in second place at RoboCup-2001 in the Sony Legged League. This is our fourth year competiting in this league. We used a simplified architecture this year. Our vision, localization, and behaviors ran synchronously with the camera input while our motions remained an asynchronous process. We focused on reducing latency and increasing timeliness and completeness of environment modelling for this year. We used more structured software engineering this year than in previous years and found this very useful[4].

This paper describes a highly modular hierarchical behavior-based control system for robots. Key features of the architecture include: easy addition/removal of behaviors, easy addition of specialized behaviors, easy to program hierarchical structure, and ability to execute nonconflicting behaviors in parallel. The architecture uses a unique reward based combinator to arbitrate amongst competing behaviors such as to maximize reward. This behavior system was successfully used in our Sony Legged League entry in RoboCup 2000 where we came in third losing only a single game.

This paper describes a motion system for a quadruped robot that performs smooth transitions over requested body trajectories. It extends the generality of path based approaches by introducing geometric primitives that guarantee smoothness while decreasing (and in some cases entirely removing) constraints on when and what types of parameter transitions can be made. The overall motion system for the autonomous Sony legged robot that served as our test-bedis also described. This motion system served as a component in our entry in the RoboCup-2000 world robotic soccer championship, in which we placed third, losing only a single game.

This paper describes a completely implemented, fully autonomous software system for soccer playing quadruped ro\-bots. The system includes real-time color vision, probabilistic localization, quadruped locomotion/motion, and a hierarchical behavior system. Each component was based on well tested algorithms and approaches from other domains. Our design exposed strengths and weaknesses in each component, and led to improvements and extensions that made them more capable in general, as well as better suited for our testing domain. Integrating the components revealed design assumptions that were violated. We describe the problems that arose and how we addressed them.The integrated system was then used at the annual Robo\-Cup robotic soccer competition where we placed third, losing only a single game. We reflect on how our system addressed its goals and what was learned through implementation and testing on real robots.

This is a description of Carnegie Mellon University’s entry in the Sony legged league of RoboCup 2000. See our web page for more details [4]. The main components of our system are: vision, localization, behaviors(including a basic world model), and motions. The main changes for this year are: basic world model, new behaviors/behavior architecture, walking and kicking motions. We placed third in the competition, losing only to the first place team.

In this paper, we report new work illustrating the integration of information agents, such as Web agents, with planning, and execution monitoring of multiple physical agents. Information agents, external to the planner and to the execution agents, specify the planning mission. The planner generates a plan which is then executed. Web agents again are requested to monitor the planrelevant features of the world. Replanning must occur dynamically when a failure is encountered. These techniques are combined in the CMUExpress architecture. The CMUExpress architecture demonstrates a solution to the integration of planning with real information and execution agents an Interaction Manager, that effectively maintains necessary links of communication and monitoring between the different sets of agents. We consider in particular multiple execution agents that must react to unforseen events while operating in the real world. Hence, in contrast to most information tasks where the world may not change while a query is being processed, we assume that the real world changes while we are solving the problem. In approaching such non-trivial real world problems, we recognize that desirable universal planning solutions may be impossible to reach as the real world is impossible to model completely. Replanning is inevitable, even to support probabilistic and conditional planning. We provide an approach to replanning that allows for the incorporation of guidance to minimally disturb the plan to be refined.

Planning actions for real robots in dynamic and uncertain environments is a challenging problem. It is not viable to use a complete model of the world; it is most appropriate to achieve goals and handle uncertainty by integrating deliberation and behavior-based reactive planning. We successfully developed a system integrating perception and action for the RoboCup-99 Sony legged robot league. The quadruped legged robots are fully autonomous and thus must have onboard vision, localization and action selection. We briefly present our perception algorithm that automatically classifies and tracks colored blobs in real time. We then briefly introduce our Sensor Resetting Localization (SRL) algorithm which is an extension of Monte Carlo Localization. Vision and localization provide the state input for action selection. Our robust and sensible behavior scheme handles dynamic changes in information accuracy. We developed a utility-based system for using and acquiring location information. Finally, we have devised several special built-in plans to deal with times when urgent action is needed and the robot cannot afford to collect accurate location information. We present results using t he real robots, which demonstrate the success of our approach. Our team of Sony quadruped legged robots, CMTrio-99, won all but one of its games in RoboCup-99, and was awarded third place in the competition.

We present a new localization algorithm, called sensor resetting localization, which is an extension of Monte Carlo localization. The algorithm adds sensor based re-sampling to Monte Carlo localization when the robot is lost. Sensor resetting localization (SRL) is robust to modelling errors including unmodelled movements and systematic errors. It can be used in real time on systems with limited computational power. The algorithm has been successfully used on autonomous legged robots in the Sony legged league of the robotic soccer competition RoboCup'99. We present results from the real robots demonstrating the success of the algorithm and results from simulation comparing SRL to Monte Carlo localization

As processor architectures increase their reliance on speculative parallel execution of sequential programs, the importance of not only what instructions to execute, but also how to resolve data dependences has increased. Data dependences present a major hurdle to the amount of instruction-level parallelism that can be exploited. Data value prediction is a technique that bypasses these dependences by speculating on the outcomes of producer instructions, allowing consumer instructions to execute in parallel. The goal of our project is to explore the application of genetic algorithms (GAs) to the design of value prediction hardware.

The CMPACK’04 team follows our past teams CMPack’03, CMPack’02, CMPack’01, CMPack’00, CMTrio’99, and CMTrio’98 [7, 4, 11, 10]. In our research this year, we explored the capabilities of the new ERS-7 robotic platform; developed new ways of modeling the world, such as using accelerometer data to detect when the robot is entangled; and maintained our focus on robust behaviors and cooperation. In this paper, we will focus in detail on several recent components of the team. First, we will describe our local model, which traces radial lines across vision frames to efficiently construct a model of the environment [5]. Next, we will describe how accelerometer data can provide a rich source of information about the world; specifically, we will show how to use accelerometer data to detect when the robot is entangled, and, as an aside, how this same algorithm can be used to identify the surface under the robot [9]. Finally, we will describe how the robot learns the effects of its different kicks by experimentation and how it then applies this knowledge to select the appropriate kick from our behaviors [3].