Pi Robot's Goals and Aspirations

Last updated: 02/15/10

Pi Robot was designed to provide a test bed for a number of ideas in robotics, cognitive science and artificial intelligence. The overall goal of the project is to build a robot that can move about autonomously in both indoor and outdoor environments while interacting with objects and people in the following basic ways:

• Avoid obstacles and drop offs (like stairs). No point in breaking the robot on its first trial run! (Although I have a funny story about that...)

• Navigate between any two locations in a known environment, or generate an internal map of a new environment through exploration.

• Orient and attend to interesting events such as moving objects, sounds and specific targets. (e.g. "Watch the moving baby while ignoring everything else.")

• Be able to pick up, carry and otherwise manipulate objects in the environment.

• Recognize basic objects (balls, boxes, bushes, beer cans), animals (dogs, cats) and people.

• Recognize familiar individuals by name.

• Follow a moving person in an appropriate way; e.g. three feet to one side while facing forward.

• Respond to basic verbal instructions using speech recognition and text-to-speech.

• Understand basic conceptual relationships among object categories (e.g. semantic network).

• Learn from experience.

Over the past several years, I have made some progress toward several of these goals. However, one learns very quickly how complex even the simplest tasks can be to program. Indeed, one of the strengths of robotics is that it is nearly impossible to fake an understanding of a particular perceptual or behavioral process. For example, if your obstacle avoidance algorithm has a flaw, the robot will run into things. By the same token, it is very gratifying when you get an algorithm "right": suddenly the robot is able to do something new right before your eyes.

In the following sections, I briefly summarize the approaches I have taken toward the goals listed above. As more progress is made, I will post updates on the Pi Robot blog.

Overall Programming Architecture

Pi Robot's control algorithms are part of a multi-threaded behavioral program. The multi-threaded nature of the program allows more than one action to be executed in parallel. For example, if you watch the video of Pi Robot tracking the yellow balloon with both his head and arms, two behaviors are actually being executed simultaneously in two separate threads: the head tracking routine that moves the servos to keep the balloon centered in the visual field, and the arm tracking routine that keeps the hands pointed at the current location of the balloon.

The two key components of Pi's programming are a pair of constructs called services and behaviors. (These are implemented as classes in programming code.) Services are wrappers around the raw data provided by Pi's various sensors such as sonar, infrared, video, etc., while behaviors are wrappers around motor actions involving servos and drive motors. A fuller explanation is given in the next two sections.

Services

An service for a given sensor can be queried or polled for a number of different values or functions involving that sensor's data. For example, the service for a particular sonar sensor can report the following information about the data stream coming from that sensor:

• Current distance reading.

• Running average of the distance reading.

• Current speed and acceleration; i.e. the rate of change of the reading, and the rate of change of the rate of change. In the case of distance sensors like sonar and infrared, these translate directly into the physical speed and acceleration of an object or the robot.

• Current "trend" of the reading, either -1, 0 or +1 for decreasing, staying the same or increasing.

• ETA or estimated time of arrival. For distance sensors, this can provide the estimated time-to-collision between the sensor and a moving object. This value is used in the video called "Pi Robot Plays Catch" to predict the moment at which the arms should start moving toward each other to intercept the approaching ball.

• Short term memory: successive sensor readings are stored in a variable length buffer so that earlier values can be retrieved if needed. This is particularly useful during learning. For example, if the goal is to avoid dead ends while exploring a new environment, then it helps to know the series of actions and sensor readings that were experienced just before getting into a dead end in the first place.

• A status message indicating the current state of the sensor reading. For example, one can set a threshold value such that the expert returns an "Alert" message when that value is exceeded.

• Other sensor calculations can be added to the expert class at any time without affecting existing code.

Services are similar to another programming construct common in robotics and AI research called agents. The notion of cognitive agents was popularized by Marvin Minsky in his book The Society of Mind which grew out of his research with Seymour Papert in the early 1970s.

Behaviors

A behavior is a single action thread that controls some subset of Pi Robot's servos or drive wheels and moves them in a certain way. The behavior can be a fixed sequence of motor movements—also called "ballistic"—or it can depend on the output of one or more services in a continual manner—such actions are also called "closed loop". An example of a ballistic behavior is the moving of the arms into a fixed position, such as up and open, regardless of the starting position or sensor input. Parameters to this action might include the speed at which the movement takes place. An example of a closed loop behavior is the moving of the two head servos (pan and tilt) to track a visual object. In this case, the motor commands are precisely dependent on the output of the vision service which reports back the x-y coordinates of the target object in the video stream.

If two behaviors do not compete for the same set of servos or drive wheels, then they can be executed in parallel in their own threads. Returning to the video of Pi Robot tracking the yellow balloon with both his head and arms, two behaviors are running simultaneously—the head tracking behavior and the arm behavior action. Since the two behaviors use different sets of servos, they can run at the same time without conflicts.

Behaviors are assigned a priority and one action can block another if it has higher priority. This is accomplished by setting a status parameter on an behavior. If a behavior's status is set to "Blocked", then the thread will be suspended for one action cycle. Once the "Blocked" status is cleared, the behavior is free to continue running. An example of this can be seen in the video Pi Robot and his Yellow Balloon. There are six threads working together in this video: visual search, head tracking, orienting, approach, pickup and throw. When Pi rotates to orient toward the ball, the approach thread is blocked. The reason is that both orienting and approaching require use of the drive wheels. Therefore, Pi must finish orienting toward the balloon before he can start moving toward it.

To be continued...