Moravec’s paradox: the hidden work of picking up a cup

Picking up a cup hides a long chain of decisions. You locate it, reach around nearby objects, make contact, and adjust your grip. A robot needs ways to solve those problems too.

Moravec’s paradox names a surprising observation from AI history. Some tasks that demand conscious human effort proved easier to automate than perception and movement that feel routine. The cup makes that hidden work visible.

What Moravec’s paradox means

In his 2003 essay “Robots, After All,” Hans Moravec revisits an early contrast. Programs made progress on narrow reasoning and game-playing tasks while robots struggled to find and move simple objects. Human judgments of difficulty gave researchers a poor guide to engineering effort.

Moravec proposed an evolutionary explanation: perception and movement draw on abilities shaped over a long history. Much of that work happens outside conscious thought. Formal calculation, which can feel laborious to us, follows explicit rules that suit computers.

Each task still needs its own assessment. Difficulty depends on the task, machine, environment, and standard for success. The historical observation cannot establish a universal ranking of reasoning and movement.

Follow one cup through the task

Imagine a robot arm with a camera and a two-finger gripper. Its goal is to lift an empty cup from a table and place it on a marked spot. The cup’s position can change between attempts.

We can split the task into five questions:

• Find: Where is the cup, and how does it face?
• Reach: Can the arm approach without an unwanted collision?
• Grasp: Where should the fingers touch, and with what force?
• Check: Did the cup move with the gripper?
• Recover: What should change after a failed attempt?

Illustrated task breakdown

One grasp, repeated observations

Each step asks a question about the same cup.

These questions form a teaching breakdown. A working system can sense and adjust throughout the motion. It may answer some questions with learned models and others with geometry or control rules.

Find the cup and estimate its pose

A label saying “cup” answers which kind of object the camera sees. Reaching also needs its location and orientation, together called its pose. The robot must relate that estimate to the arm’s coordinates.

The camera sees only the surfaces visible from its viewpoint. Another object might hide part of the cup. A useful estimate must account for what the sensor cannot resolve.

Coordinate frames explain how the same location gets different numbers in camera and robot coordinates. Camera calibration establishes the relationship between those frames. An error there can send an accurately controlled gripper to the wrong place.

The machine-learning fields map connects this perception problem to computer vision. Learning can help interpret an image; the rest of the task still needs a way to use that interpretation.

Make room for the gripper

A reachable cup can still be awkward to approach. The arm must fit around the table, the cup’s handle, and any nearby objects. Its final grasp also needs a suitable gripper orientation.

Russ Tedrake’s MIT pick-and-place notes begin with a known object pose and still need geometry and kinematics. They introduce an approach pose before the grasp. Even a perfect pose estimate leaves motion decisions to solve.

Try a small, invented clearance calculation. Viewed from above, a rigid cup body is 70 mm wide. Two parallel fingers have an 80 mm gap between their inner faces.

With both centers aligned, the clearance on each side is:

(80 − 70) / 2 = 5 mm

A 4 mm sideways position error leaves just 1 mm on the tighter side. That calculation checks one cross-section. It leaves out the handle, finger thickness, approach path, and deformation.

Robot workspaces explain reachability. Collision checking adds the space occupied by the moving bodies along the route.

Turn contact into a useful grasp

Closing the fingers changes the problem. Forces now pass through contacts between the fingers and cup. The grasp must support the load as the arm lifts it.

For a simple side grasp, friction can resist the cup’s downward slip. In the approximate Coulomb model, the available tangential force depends on the normal squeeze force and friction coefficient. Lynch and Park’s friction lesson explains that limit and its assumptions.

More squeeze can increase the friction available in that model. A real gripper also has force limits, and a thin cup can deform. A useful command must account for those limits and the load it needs to support.

Wrenches describe contact forces and their turning effects. Robot statics connects a load at the tool to the joint torques needed to hold it.

Check whether the cup moved

The gripper followed its command. Did the cup follow the gripper?

A camera might track the cup’s height. Tactile sensors might detect contact or slip. Each signal answers a different question, and each can contain errors.

Feedback uses measurements to change the next command. If the cup shifts, a controller may adjust within its allowed force and motion limits. If the evidence is insufficient, the system needs a way to pause or seek another observation.

The feedback-control lesson makes this correction visible in a simpler joint-speed model. It also shows why a biased measurement can produce a wrong correction. More feedback alone cannot guarantee a successful grasp.

Recover when the attempt fails

Suppose the fingers close, but the cup stays on the table. Repeating the same motion with the same estimate gives no reason to expect a better result. The robot needs evidence that changes its next attempt.

For this task, a recovery could involve:

• Opening the gripper and retreating along a checked path.
• Taking another view to update the cup’s pose.
• Choosing a different approach or contact location.
• Stopping when the available observations cannot support another attempt.

Those actions need their own checks. A retreat path might become blocked after the cup moves. A recovery rule must account for the scene that exists now.

Read a demonstration’s limits

A robot can succeed at the cup task. To interpret that success, ask what the test actually covered.

• Did every attempt use the same cup and lighting?
• Who chose the starting pose and grasp location?
• How many attempts succeeded, out of how many total attempts?
• Did the system detect failures and recover without help?
• What changed in the trials that followed?

A fixed setup can support a useful application. Claims about unfamiliar cups or changing conditions need trials that cover those changes. A successful demonstration supplies evidence for the conditions it tested.

Moravec’s paradox is useful here as a prompt to inspect hidden assumptions. Apparent ease for a human leaves the robot’s engineering requirements unanswered.

These questions also help you scope a portfolio project. Pick one part of the task, record its conditions, and explain a failure you can reproduce. The robotics career guide's project examples connect that evidence to controls, software, learning, and tooling roles.

Try it yourself

Exercise 1: spend the clearance

Keep the 80 mm finger gap and 70 mm rigid cup body. The cup’s center lies 6 mm to the right of the gap’s center. Does it fit between the fingers in this one-dimensional model, and by how much?

Check the clearance answer

The inner finger faces sit at −40 mm and +40 mm. The cup occupies the interval from 6 − 35 = −29 mm to 6 + 35 = 41 mm. Its right edge extends 1 mm past the right inner face, so it does not fit.

Positive clearance on both sides requires a center error smaller than 5 mm in magnitude. At exactly 5 mm, the cup touches one inner face. This check alone says nothing about a secure grasp or a clear approach path.

Exercise 2: choose evidence for recovery

After a commanded lift, the gripper reports that its fingers closed fully. A clear camera view shows the cup still resting on the table. Should the system mark the lift complete, and what should it check before another attempt?

Check the recovery answer

The lift failed in this scenario: the cup stayed on the table. Finger position reports the gripper’s state, and the camera supplies evidence about the object’s motion.

Before retrying, the system should update the cup’s pose and check a retreat and approach path. It also needs to decide whether the proposed contact location still makes sense. If it cannot establish those conditions, stopping is a valid outcome.

The physical-AI overview follows a mobile robot through observation, estimation, action selection, and feedback. Use that system view to locate the hidden work behind a physical task and identify where learned models can help.

Sources and further study

• Hans Moravec, “Robots, After All” (2003): an author-hosted account of the historical contrast and his evolutionary explanation. Its forecasts belong to their historical context.
• Russ Tedrake, “Basic Pick and Place”: a worked treatment of poses, approach motions, and robot kinematics, initially assuming perfect object-pose information.
• Kevin Lynch and Frank Park, “Friction”: the contact-force model behind the simple grasp discussion.

Choose one part of the cup task to examine more closely. Start with robot workspaces and test whether a reachable point also permits the tool orientation you need.