Shakey the robot: primary technical report

From 1966 through 1972, the Artificial Intelligence Center at SRI conducted research on a mobile robot system nicknamed "Shakey." This research was sponsored by the Advanced Research Projects Agency under a succession of contracts with the Rome Air Development Center, the National Aeronautics and Space Administration, and the Army Research Office.

It is the objective of this program to develop concepts and techniques in artificial intelligence enabling an automaton to function independently in realistic environments. These concepts shall be demonstrated by means of a breadboard, mobile vehicle containing visual, tactile, and acoustic sensors, signal processing and pattern-recognition equipment, and computer programming. Primary goals shall be the solution of incompletely specified problems (requiring creation of intermediate strategies and goals) and improvement of performance with training experience.

Second, we decided to minimize hardware complexities whenever possible to allow us to focus primary attention on the problem of directing the automaton's actions and planning by means of a hierarchy of computer programs. For this project the mechanical engineering problems of building a robot with articulated limbs and delicate grasping abilities are irrelevant. One can face very tough problems in artificial intelligence directly in attempting to write computer programs to control even a very simple vehicle.

Shakey used a planning system called STRIPS (an acronym based on STanford Research Institute Problem Solver) to chain together ILAs that would accomplish specific goals. STRIPS was one of the important early problem-solving systems.

The robot system keeps track of where the robot vehicle is and stores its other knowledge of the world in a model composed of well-formed formulas in the predicate calculus. Tasks are given to the system in the form of predicate calculus formulas. To direct the robot to go to room R2, we pose for it the goal formula INROOM(ROBOT,R2). The planning system, STRIPS, then attempts to find a sequence of primitive actions that would change the world in such a way that the goal formula is true in the correspondingly changed model.

Within this basic framework STRIPS operates in a GPS-like manner. First, it tries to establish that a goal formula is satisfied by a model. If the goal formula cannot be proved, STRIPS selects a "relevant" operator that is likely to produce a model in which the goal formula is "more nearly" satisfied. In order to apply a selected operator, the precondition formula of that operator must of course be satisfied: this precondition becomes a new subgoal and the process is repeated.

The robot vehicle itself is propelled by two stepping motors independently driving a wheel on either side of the vehicle. It carries a vidicon television camera and optical range-finder in a movable "head." Control logic on board the vehicle routes commands from the computer to the appropriate action sites on the vehicle. Besides the television camera and range-finder sensors, several "cat-whisker" touch-sensors are attached to the vehicle's perimeter. These touch sensors enable the vehicle to know when it bumps into something. Commands from the computer to the vehicle and information from the vehicle to the computer are sent over two special radio links, one for narrow-band telemetering and one for transmission of the TV video from the vehicle to the computer.

The PDP-10 system has 192K words of 36-bit memory. The Bryant drum is a high-speed autolift drum which has a 1.5-million-word capacity. The TV A/D converter is an SRI-designed and -built device. It handles data from the robot TV camera at a rate of one word every 1.5 microseconds. It is capable of processing either 120x120 or 240x240 pictures with 32 levels of gray scale.

The current Automaton is restricted by its method of locomotion to move only on nearly flat surfaces. The first visual subsystem was designed to specialize in the planar-surfaced environment of our laboratory and office building. The objects in this environment are specially constructed rectangular parallelepipeds and wedges.

Perhaps the most difficult problem that confronts the designer of ILAs is the problem of detecting and recovering from errors. Sometimes errors are detected automatically, as when an interrupt from a touch sensor indicates the presence of an unexpected obstacle. Other times it is necessary to make explicit checks, such as checking to be sure that a door is open before moving through it. When an error is detected, the problem of recovery arises. This problem can be very difficult, and is one aspect that distinguishes work in robotry from other work in artificial intelligence.

To circumvent this problem, we have chosen to have the subroutines communicate through the model. With a few special exceptions, neither answers nor error messages are explicitly returned by subroutines. Instead, each routine uses the information it gains to update the model. It is the responsibility of the calling routine to check the model to be sure that conditions are correct before taking the next step in a sequence of actions.