Evolutionary Blueprint (3). Why loading a dishwasher is still a hard problem for robots

Series: Evolutionary Blueprint of AI, article 3. A robot can describe a kitchen and still struggle with a wet plate. Learn why physical tasks are difficult, where robotics research has made progress, and how to assess an automation proposal.

Evolutionary Blueprint (3). Why loading a dishwasher is still a hard problem for robots
Illustration generated with ChatGPT. The gap between 2 worlds.

Who should read it?

Technology and business leaders, product teams and curious non-specialists.

Metapher

The gap between 2 worlds.

Intro

Imagine a company testing a robot in its office kitchen. It asks the machine to load the dishwasher. The robot identifies a plate on the counter, reaches for it and closes its gripper. The plate is wet, slips and comes to rest against a mug. The robot now has a different scene from the one it planned for, and a person has to intervene.

This is a fictional example, not a report of a particular robot. It illustrates a familiar difference:

  • Describing a sequence of actions is not the same as carrying it out

A working dishwasher loader must handle different objects and positions, notice when an action fails and avoid harming people or damaging equipment. Some robots can perform pieces of this work under specified conditions. The difficult question is how reliably an entire system handles the conditions it will actually encounter.

We use Moravec's paradox to examine that difference. The lesson for a business is not to avoid physical automation. It is to specify the task and its limits before treating a demonstration as an operating plan.

What Moravec's paradox does and does not say

Robotics researcher Hans Moravec observed a mismatch between human intuitions about difficulty and the abilities of computers:

  • Tasks associated with adult reasoning, such as board games, had proved more tractable for machines than some ordinary abilities involving perception and movement

His own research on a mobile robot illustrates the challenge:

  • The system planned routes around obstacles using camera images, but moved slowly and was reliable only for short runs in the reported experiments.

The observation is often called Moravec's paradox. It is a useful heuristic, not a theorem that proves reasoning is cheap or movement is always expensive. The cost depends on the task, the hardware, the available data and the reliability required. A tightly controlled factory robot can repeat a movement very well, while a digital model may struggle with a difficult proof or consume substantial computing resources. Technologies change, too, so examples that once seemed impossible may later become achievable.

The contrast also does not mean digital errors are harmless. A language model may write an incorrect sentence, but an assistant connected to a payment system can initiate a costly action. Physical robots add direct constraints such as contact forces and proximity to people. Both digital and physical systems need controls matched to the harm they could cause.

A plate presents several problems at once

A person may put a plate in a rack without thinking about each movement. A robot system must deal with at least the following uncertainties:

  • Seeing the scene.
    • Another dish may hide the edge of the plate. Reflections, shadows or a transparent glass can make some sensors less dependable.
  • Choosing a grip.
    • The plate may be heavy, wet, chipped or unusually shaped. A grip strong enough to hold it should not damage it.
  • Moving through space.
    • The arm needs clearance around the counter, the open door and the other dishes. Its gripper and joints limit which approaches are possible.
  • Checking the result.
    • A command to close the gripper does not establish that the plate is secure. If it slips or the rack has moved, the system needs a way to detect the problem and stop, retry or ask for help.

These are interacting problems rather than independent boxes to tick. A small error in estimating the plate's position can affect the grip. A failed grip changes the scene for the next attempt. A robot need not explicitly calculate every force from first principles. It might use a learned policy, a physical model, feedback control or a combination. Whatever method it uses, the result must be tested under the conditions that matter.

In Dex-Net 2.0, Jeff Mahler and colleagues trained a grasp planner using synthetic data and tested it on physical objects, including objects not used for training. They also described limitations in their sensors and grasping model. This is evidence that simulated examples can support real grasping in a particular setup, not proof that the same system can load an arbitrary dishwasher or handle every transparent, flexible or slippery item.

Simulation helps, but the kitchen gets the final test

Collecting robot data can be slow. A physical attempt takes time, equipment and sometimes a person to reset the scene. Simulation lets a team generate and vary examples without handling a real plate each time. It can also expose a controller to situations that would be inconvenient or unsafe to recreate during early development.

The limitation is the simulation-to-real gap:

  • A simulation is an approximation of the sensors, objects and environment

A simulated gripper may not reproduce the wear of a real gripper. A simulated plate may behave differently when wet. A camera may see a scene differently under actual lighting. These differences can cause a controller that works in simulation to fail in deployment. They do not make simulation useless or guarantee failure.

Anthony Brohan and colleagues' RT-1 study tested a robot model trained on real demonstrations across multiple manipulation tasks, including placing a water bottle upright. In a separate experiment, the researchers added simulated training data for objects absent from real training and then tested physical robots on real versions of those objects. Results improved on the tested pick and move-to cases compared with training on real data alone. The study involved specific robots, tasks and evaluation conditions. It does not establish reliability for every kitchen or a new customer's site.

For a buyer, that distinction changes the acceptance test. Ask which objects, locations and conditions were held out from development. Then repeat trials at the intended site with its actual racks, lighting and workflow. A clean laboratory video can be useful evidence of possibility, but it is not an estimate of production performance.

What a body can contribute to learning

Bennett's evolutionary perspective draws attention to abilities that humans rarely notice because they are so practiced. Animals sense their surroundings and act within them; physical skills depend on bodies, environments and feedback. That history is a reason to take movement seriously, not a blueprint proving that an artificial system must replay biological evolution.

Embodied cognition is a family of research perspectives on how bodies and their interactions with environments contribute to cognition. As Lawrence Shapiro and Shannon Spaulding explain, those perspectives differ, and their stronger philosophical claims are debated. One practical idea is easy to illustrate:

  • A robot that can touch an object and observe whether it slips can use feedback unavailable in a single still image.

That does not establish that a physical body is necessary for every form of understanding, or that a robot automatically understands because it has one. Systems can learn about objects through language, images, video, simulation and interaction. Each source can provide useful information and has limits. A description of a cup can convey that it holds liquid; a physical trial can reveal whether a particular gripper can lift that cup. Neither alone guarantees success in the next kitchen.

Turn the demonstration into an operating decision

Suppose the fictional company wants to automate its office kitchen. Before buying a general-purpose robot, the operations team should compare alternatives:

  • Change the rack or dish selection, use a more specialized machine, keep the task manual, or pilot a robot with a deliberately narrow set of dishes.

A simpler environment may make automation more practical, but standardizing dishes and changing staff routines also have costs.

For the pilot, distinguish a successful placement from a completed load. A placement succeeds when the robot puts an approved plate into its assigned rack slot without damage and leaves it stable enough for the wash cycle. A load is complete only when every item in the agreed load is correctly placed and the dishwasher is ready to run. State which items, initial positions and environmental conditions are in scope, and what should happen when an unfamiliar item appears. A safe stop and handoff can be the correct outcome for an exception. It should not be counted as a completed load.

Test more than successful grasps. Include full attempts with partially blocked plates, wet surfaces, mixed dishes and a rack that is not in its expected position. Record the conditions, the number of attempts, completed loads, damage, missed detections, stops and human interventions. Report results by condition as well as overall:

  • An average can conceal that the robot fails on the glassware used every morning

Compare the process with the current manual workflow, including the time staff spend preparing, resetting and recovering the system.

Before the pilot, the process owner and safety team should set acceptance criteria appropriate to the site and complete a task-specific risk assessment. Plan safeguards for people working nearby, emergency stops, maintenance, fault recovery and who is authorized to restart the system. A learned model's confidence or a successful demonstration is not a substitute for those safeguards. Use a controlled test area first, and do not expand the scope simply because a handful of trials worked.

Finally, compare cost and staff time per completed load with the current manual baseline over comparable loads and conditions. Include equipment, integration, maintenance, energy, training, supervision, downtime and damage, as well as preparation, resets and recovery for both workflows. State how fixed costs are allocated and count unsuccessful attempts in the total costs and time, even though they add no completed load to the denominator. If staff must sort every load and frequently rescue the robot, a faster individual pick may not produce a faster or cheaper kitchen operation. The same logic applies in a warehouse or a factory, although the objects, safety requirements and economics will differ.

A successful pick is not a successful operation

Moravec's paradox is a prompt to question human intuitions about what should be easy for a machine. It is not a universal ranking of computational difficulty. Physical automation can already be useful in bounded settings, and research has shown progress on learning and grasping across varied tasks. Reliability still has to be established for the actual equipment, objects and people involved. Define the operating conditions, test exceptions and count the full cost before turning a robot demo into a business commitment.

Previous

Article 2: Your brain predicts the world

Next

Article 4: Reward is not the whole lesson