Shakey the robot: Shakey enters the world

The event and the narrow claim

In March 1966, a group at the Stanford Research Institute (now SRI International) in Menlo Park began building a wheeled robot that would be nicknamed Shakey. The documented claim is precise and modest: over the following six years, SRI's Artificial Intelligence Center built a mobile machine that could perceive a simplified indoor world, maintain a symbolic model of it, plan a sequence of actions toward a goal, and then carry that plan out while checking for failures. SRI's own account of the project describes it as the first mobile robot able to reason about its own actions rather than replay a fixed script.

The broader judgment (that this is where artificial intelligence stepped out of the puzzle box and onto a floor) deserves to be stated separately and carefully. The date attached to this timeline entry, 1966, marks the start of a program, not a finished invention delivered in a day. The DARPA innovation record notes that the work grew from a 1964 SRI proposal for research in "Intelligent Automata" and was funded by the agency then called ARPA. What makes the event matter is not that one machine walked and thought at the same moment, but that a team committed to integrating perception, world modeling, planning, and motion in a single system, and in the attempt produced algorithms and an architecture that outlived the robot itself.

From fixed motion to reasoned action

To see the constraint Shakey addressed, contrast it with the machines that already existed. Industrial automata of the early 1960s, most famously the Unimate arm installed on a General Motors line in 1961, could move with great precision, but they repeated preprogrammed trajectories and had no model of what they were doing or why. They were physical without being deliberative. In the other direction, the reasoning programs of the era ran on symbols alone, in worlds that existed only inside the computer.

Shakey's contribution was to insist on both halves at once. A robot that reasons about its actions must convert noisy sensor data into a representation it can think about, choose actions whose effects it can predict, and detect when the physical world fails to match its expectations. Each of those problems, taken alone, was already hard; wiring them into one control loop that closed from camera to plan to wheels and back was the research goal. As Nils Nilsson later put it in his retrospective, Shakey: From Conception to History, the project was deliberately designed to force the integration of subfields that were otherwise studied in isolation.

The machine on the floor and the mind in the room

Physically, Shakey was an unglamorous stack of hardware on a wheeled base, roughly the height of a person, so named because it shook noticeably when it started and stopped. It carried a television camera, an optical rangefinder, and touch-sensitive "cat whisker" bump detectors, and it drove on wheels turned by stepping motors. What it did not carry was its brain. As the Computer History Museum's exhibit note records, Shakey communicated over a two-way radio link with large remote computers that did the actual thinking; a proposed manipulator arm was designed but never built, so the robot's only way to move an object was to push it.

That division of labor is easy to underrate. The computation ran on a DEC PDP-10 mainframe, with a smaller PDP-15 handling lower-level device control, machines that filled a room and were shared among many users. Shakey was, in effect, a mobile set of eyes and wheels tethered wirelessly to a time-shared computer. This is a reminder that the achievement was architectural and algorithmic rather than a triumph of onboard miniaturization. The intelligence lived in software running elsewhere, and the robot was the instrument through which that software touched the world.

The sense, model, plan, act loop

Shakey's control software was organized as a hierarchy, and understanding it clarifies what "reasoning about actions" meant in practice. At the bottom were low-level actions: primitive commands such as rolling a set distance or turning through an angle, expressed in the robot's own motor terms. Above these sat intermediate-level actions that combined primitives into useful routines, for example navigating to a named location or pushing an object while correcting course. At the top sat a planner that reasoned over a symbolic model of the world, deciding which intermediate actions to invoke and in what order.

This layering is the ancestor of what later roboticists named the sense, plan, act paradigm. The world model was a collection of logical statements about rooms, doorways, boxes, and the robot's own position. Perception updated those statements; planning consumed them to produce a course of action; execution turned the plan into motion; and the outcome was sensed again, closing the loop. Nilsson's book-length history of the field, The Quest for Artificial Intelligence, treats Shakey as the canonical early demonstration that these stages could be joined in one working system rather than merely described.

STRIPS: actions as preconditions, add lists, and delete lists

The planning layer is where Shakey left its deepest theoretical mark. Richard Fikes and Nils Nilsson built a planner they called STRIPS, the Stanford Research Institute Problem Solver, described in their 1971 paper STRIPS: A New Approach to the Application of Theorem Proving to Problem Solving. STRIPS represented the world as a set of first-order logic formulas and represented each possible action as an operator with three parts: a precondition that must hold for the action to apply, an add list of facts the action makes true, and a delete list of facts it makes false.

That representation was the crucial simplification. Earlier attempts had tried to prove, from scratch, that a goal state was reachable, an approach that drowned in the bookkeeping of which facts change and which persist (the so-called frame problem). By declaring exactly what each action adds and deletes and leaving everything else untouched, STRIPS made planning tractable enough to run. It searched by means-ends analysis, repeatedly asking which operator would reduce the difference between the current world and the goal, a technique inherited from earlier general reasoning research. A command like "push the box into the next room" could then be decomposed automatically into the subgoals of reaching the box, orienting behind it, and pushing it through the doorway.

Executing a plan and noticing when it breaks

A plan that assumes the world behaves perfectly is useless to a robot that shakes, slips, and mis-senses. Shakey's answer was a plan-execution layer, PLANEX, built on a data structure called the triangle table. As described in accounts of the project's plan execution and monitoring work, a triangle table recorded each action of a plan together with the conditions it required and the effects it was expected to produce.

This gave Shakey two capabilities that a flat script could not. First, during execution the system could compare the world it actually sensed against the outcome the table predicted, and if they diverged, it could detect the failure rather than blindly continuing. Second, on failure it could replan from the current state while reusing the still-valid portions of the original plan, instead of starting over. The triangle table also let a successful plan be generalized and stored as a reusable macro operator, an early form of learning from experience. Execution monitoring and recovery, taken for granted in modern autonomy, were treated here as first-class problems.

Seeing a room made of lines

Between the camera and the logical world model lay computer vision, and here too the project produced lasting mathematics. Shakey's world was engineered to be seeable: rooms had clean baseboards and objects were large, flat-sided boxes and wedges under controlled lighting. The vision system's job was to find the edges of those objects and the boundaries of the floor from a grainy television image and to translate them into positions in the model.

To find straight lines robustly in noisy images, Richard Duda and Peter Hart formalized and popularized a method now known everywhere in image processing, published as Use of the Hough Transformation to Detect Lines and Curves in Pictures in 1972. Their key move was to parameterize a line by an angle and a distance rather than by slope and intercept, which avoided the infinities that plague vertical lines and turned line detection into a voting procedure in parameter space. The Hough transform came directly out of the perception needs of a physical robot, and it remains a standard tool in computer vision decades later, a concrete example of practical robotics driving general theory.

What Shakey actually did, and how slowly

The honest evidence is a set of demonstrations in a controlled space, not open-world competence. Shakey operated in a suite of connected rooms containing walls, doorways, and a handful of boxes and wedge-shaped ramps. A typical and frequently cited task began with an English-language instruction to push a particular block off a raised platform. Shakey had no way to reach the platform directly, so its planner inferred an unstated subgoal: locate a ramp elsewhere in the environment, push the ramp against the platform, roll up it, and push the target block off. The point of the demonstration was that the sequence was not preprogrammed; it was derived by the planner from a description of the goal and the available operators.

The performance came at a cost that is important to state plainly. Shakey was slow. A task could involve long pauses while the remote computer analyzed an image or searched for a plan, and complex missions could take hours of stop-start operation. The environment was carefully prepared to make perception feasible. The vocabulary of objects and actions was small. These are not incidental details; they are the boundary of the claim. Within that boundary the system genuinely perceived, reasoned, acted, and recovered, which was new. Outside it, Shakey could do nothing.

A* and the cost of thinking about where to go

Navigation posed its own problem: given a map of the rooms, find a good route to a destination without examining every possible path. The solution, developed by Peter Hart, Nils Nilsson, and Bertram Raphael and published in 1968 as A Formal Basis for the Heuristic Determination of Minimum Cost Paths, is the A* search algorithm. A* orders its search by the sum of the cost already spent to reach a node and a heuristic estimate of the cost remaining to the goal.

The paper's enduring result is a guarantee, not just a program: when the heuristic never overestimates the true remaining cost (a property called admissibility), A* is guaranteed to return a lowest-cost path, and among such algorithms it expands the fewest nodes. That combination of optimality and efficiency is why A* escaped the robotics lab entirely. It is now a workhorse of route planners, video-game pathfinding, and countless scheduling and search problems, arguably the single most widely used artifact the project produced. Like the Hough transform, it is a case of a general algorithm crystallizing out of a specific engineering need.

The rival vision: reasoning without a body

Shakey's approach was not the only bet of its era, and the most instructive contrast lived at MIT. There, Terry Winograd's SHRDLU, documented in his 1972 work Understanding Natural Language, carried on fluent English conversations about a "blocks world" of stacked shapes and could answer questions, take instructions, and explain its reasoning. Crucially, SHRDLU's blocks world was entirely simulated. It had no camera, no wheels, and no gap between what it believed and what was physically true.

Placing the two side by side sharpens what each demonstrated. SHRDLU showed how rich symbolic reasoning and language understanding could be when the world was clean and fully known. Shakey showed how much harder, and how different, the problem becomes once perception is imperfect and actions can fail. Both belonged to the symbolic, logic-based tradition that dominated AI at the time, and both would later be pressed by the difficulty of scaling that tradition to messy reality. The divergence between reasoning in a modeled world and acting in a physical one is a fault line that still runs through robotics research.

The downstream inheritance, and how to place the event

The clearest way to weigh Shakey is to trace what later work reused. A* became ubiquitous in navigation and search. The STRIPS action representation became the template for automated planning, evolving through successors into the Planning Domain Definition Language that planning research still uses. The sense, plan, act architecture and the practice of execution monitoring became standard reference points in autonomous robotics, cited in the lineage of planetary rovers and self-navigating vehicles. In 2004 Shakey was inducted into Carnegie Mellon's Robot Hall of Fame, and in 2017 it received an IEEE Milestone, the first robot so recognized, with the robot itself now on display at the Computer History Museum.

The cautions matter as much as the credits. Contemporary press coverage ran ahead of the science: Life magazine in 1970 called Shakey the "first electronic person," a phrase SRI itself notes even as it records the more sober engineering reality. Shakey did not perceive, plan, or act at anything like human generality; it operated in a scaffolded world at a slow pace, its intelligence borrowed from room-sized computers. The right placement in AI history is as an integration milestone rather than a single invention: the project that first insisted perception, symbolic reasoning, planning, and physical action be solved together, and that, in solving pieces of that problem, left behind tools (A*, STRIPS, the Hough transform, execution monitoring) more durable than the machine that occasioned them. That the timeline dates the event to the project's beginning is fitting, because its significance lies in the direction it set, not in a finished result.

References

• SRI International (2017). Shakey the Robot.
• SRI International (2021). 75 Years of Innovation: Shakey the Robot.
• Hart, Nilsson and Raphael (1968). A Formal Basis for the Heuristic Determination of Minimum Cost Paths.
• Fikes and Nilsson (1971). STRIPS: A New Approach to the Application of Theorem Proving to Problem Solving.
• Duda and Hart (1972). Use of the Hough Transformation to Detect Lines and Curves in Pictures.
• Kuipers, Feigenbaum, Hart and Nilsson (2017). Shakey: From Conception to History.
• Engineering and Technology History Wiki (2017). Milestones: SHAKEY, The World's First Mobile Intelligent Robot, 1972.
• DARPA (2023). Shakey the Robot.
• Computer History Museum (2011). Shakey, CHM Revolution.
• Nilsson (2010). The Quest for Artificial Intelligence.
• Winograd (1972). Understanding Natural Language.
• Robohub (2017). Shakey is first robot to receive IEEE Milestone award.