Shakey the robot: the first machine that could reason about its own actions

Shakey the Robot, SRI Technical Note 323, was published in April 1984 by Nils J. Nilsson, Director of SRI International's Artificial Intelligence Center. It is a compilation, not a single original report: Nilsson extracted and reprinted the most significant sections from twelve separate SRI technical reports produced during the original project, gathering material that had previously been scattered across relatively inaccessible individual documents into one accessible account. The document was approved for public release and is archived by the Defense Technical Information Center under accession number ADA458918.

The report states plainly that Shakey research ran from 1966 through 1972 and 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. Two complete versions of the robot were built: an initial 1969 system controlled by an SDS-940 computer, and a more powerful 1971 system using a Digital Equipment Corporation PDP-10 and PDP-15 combination.

The report is explicit about the project's founding ambition, quoting the original 1966 program goals directly: the objective was to develop concepts and techniques in artificial intelligence enabling an automaton to function independently in realistic environments, with primary goals being the solution of incompletely specified problems and improvement of performance with training experience. This is, in plain 1966 language, a description of autonomous goal-directed reasoning, not preprogrammed automation.

The technical mechanism that made this possible was STRIPS, the Stanford Research Institute Problem Solver, a planning system that let Shakey chain together intermediate-level actions to accomplish goals it had never been explicitly programmed to reach. Given a goal such as being in a particular room, STRIPS would search for a sequence of actions that would transform the robot's internal model of the world into one where that goal was true, then execute that plan, and replan if execution failed. It is a direct conceptual ancestor of the planning and reasoning systems described decades later in DARPA's Strategic Computing plan.

Shakey moved on two independently driven wheels, carried a television camera and an optical range-finder on a movable head, and used touch-sensitive whiskers on its perimeter to detect collisions. It communicated with its off-board computer over radio links, since the computing power required did not fit inside the vehicle. By the standards of the day the computing resources were substantial: the PDP-10 system used 192,000 words of 36-bit memory, and Shakey's control programs eventually occupied over 300,000 words, an amount the report itself flags as large for the period.

The robot's environment was deliberately simple: a set of rooms with painted rectangular blocks and wedges as obstacles, built specifically to be within reach of the vision technology of the time while still complex enough to expose the weaknesses of that technology. The report is candid about this limitation, noting that the first step taken toward a general visual capability was very small indeed.

This document establishes, in SRI's own technical account, that a US government-funded research program built and operated, between 1966 and 1972, a physical robot capable of building an internal symbolic model of its environment and generating novel plans of action to satisfy goals, using a planning architecture, STRIPS, that treated the world as a set of logical statements to be transformed. It establishes the specific hardware, funding sources, and technical architecture involved, all confirmed in SRI's own words.

What it does not establish is any claim to general intelligence. Shakey operated in a deliberately constrained, artificial environment of painted blocks and flat floors, and the report itself repeatedly acknowledges the narrowness of what the vision and planning systems could actually handle. Shakey is significant as a first working demonstration of the architecture, perception feeding a world model feeding a planner, not as evidence that general-purpose robot intelligence was achieved in 1972.