When Doing the Task Is Not Enough — Part 2

Part 1 of this series, “What Will Change First in Robotics?”, asked which parts of robotics will change first. This part asks what happens after a robot can do the task: how the result is judged, shaped, and lived with.

When I think about what comes after a successful robot demonstration, I often return to computer graphics. Toy Story reached theaters in 1995 as the first computer-animated feature film. It established that a complete film could be made this way. It did not establish that every character, movement, and visual effect would now be straightforward to produce. Nor was it a technical demonstration waiting for someone to add artistic judgment later. The characters already had to carry a story. Making the medium possible and making something worth watching were connected from the beginning.

A character’s arm can follow its skeleton and still deform badly at the elbow. A technically valid movement can look wrong, or be difficult for an animator to change into the movement they want. This is not a small concern added after the real problem is solved. It is part of what makes the technique useful. What interests me about this history is the relationship between a capability and the control people need over it. As more of the underlying work becomes available through tools, some demands become easier to satisfy. Other differences become easier to notice, and new uses become possible.

I expect a related change in robotics over the next few years. Not because basic reliability is about to disappear as a problem, but because some tasks will become dependable enough that success alone stops distinguishing good systems. Two robots may both hand over a cup while imposing very different effort on the person receiving it. They may differ in timing, placement, responsiveness, and how easily a mistaken instruction can be corrected. Looking ahead from June 2026, I think these differences will become part of the next capability frontier. The question is not only whether a robot can complete the task, but how people can obtain, shape, and live with the result.

Making it work does not finish the work

Skinning, the connection between a character’s skeleton and its visible surface, provides a concrete example. Direct Delta Mush, published in 2019, sought better deformation while improving computational efficiency and control over the result. It did not treat animation quality, execution cost, and ease of authoring as interchangeable. The method was valuable partly because it changed the effort required to obtain a useful deformation, not simply because it made another mesh move. I would not describe this as engineering solving functionality before artists supply quality. A new representation can make an intended correction easier and an unwanted deformation less likely. It changes both the result and the work through which judgment becomes that result. For robotics, this suggests that how a behavior is represented and edited may matter as much as the existence of a successful example.

Consider a robot folding a towel. Once it can produce a stable fold, a user may care whether the result fits a particular shelf, keeps an embroidered side visible, or matches the other towels. None of these requirements is necessarily mysterious or resistant to learning. They are simply not identical to the condition used to label an earlier trial successful. A stronger robot could learn them from examples, instructions, or corrections. But the system needs a way to preserve what is already right while changing the relevant part. Repeating the entire task until the user happens to like the result is a weak form of control.

This gives me a more specific forecast than a general increase in visual polish. Within task families where basic execution becomes reliable, robots will increasingly be compared on the cost of obtaining a particular acceptable result. Some will be easier to direct. Some will retain corrections more appropriately. Some will make their next action easier to understand. These improvements need not wait until every motor-control problem is solved. They can develop alongside reliability, and sometimes help produce it. A person who can redirect a robot early may prevent a failure that a more isolated controller would encounter later.

What a number keeps

It is tempting to describe systems as quantitative and graphics or human-robot interaction as qualitative. I understand the distinction, but I do not think it is stable enough to support a forecast. Systems researchers still choose which workload and which kind of delay matter. Graphics researchers measure execution time and deformation properties as well as asking whether an image looks right. A number does not remove interpretation from a problem. It records selected properties in a form that supports comparison. The important question is which differences it preserves for the decision we are trying to make.

A metric can be understood as a compressed description of a much richer outcome. That does not make it defective. Sometimes a small number of measurements is exactly what a decision needs. The problem appears when similar scores conceal a difference that matters. Two towel folds may both count as complete while only one fits the shelf. A better metric might capture that distinction. Or we may discover that we had been asking two different questions while treating them as one.

The LPIPS research made this issue concrete for image similarity. Its authors collected human judgments and studied how well learned visual features matched them compared with established image metrics. The results showed that aspects of perceptual similarity could be captured more effectively by learned representations. That is evidence against treating visual judgment as an undifferentiated human faculty that measurement can never reach. It is not evidence that one similarity measure also captures narrative effectiveness, artistic intention, physical correctness, or the ease of editing an asset. Those are additional questions, not merely more difficult versions of the same question.

Robotics already mixes human judgment and numerical evaluation. In the π*₀.₆ paper, task success is derived from human annotations of multiple quality indicators, and throughput counts successful executions per hour. The numerical result therefore contains prior judgments about what qualifies as success. This does not make the result less scientific. It makes the definition and consistency of those judgments part of the evidence. A graph of success rates cannot, by itself, tell us what the raters were asked to accept or which differences the label was designed to ignore.

Human-robot interaction offers an especially useful distinction. Dragan, Lee, and Srinivasa’s 2013 work separated predictable motion from motion that makes a robot’s goal easy to infer. A person who already knows the goal may expect one path, while a person trying to discover the goal may understand a different path more readily. The researchers formalized the distinction and tested its consequences. Something that could have remained a vague preference for natural movement became a more precise research question. I expect similar distinctions to become important as robots move from completing isolated tasks to making their behavior understandable during shared activity.

Learning to judge is also a technical capability

The fact that a person supplies a judgment does not mean a person must supply it on every future trial. Deep reinforcement learning from human preferences, published in 2017, used comparisons between short behavior segments to train a reward model and then improve policies in Atari and simulated robot locomotion. The system did not require a human to score every interaction. The work is older than the current generation of VLAs, but its relevance is direct. Some goals that are awkward to write as equations can still be communicated through comparisons and learned well enough to guide behavior.

Over the next few years, I expect more robot systems to use models that evaluate how a task was performed, not only models that produce actions. A critic could compare two folds, flag an awkward handover, or identify the stage where a result departed from an example. A user might correct that critic rather than demonstrate every motor movement. This would shift some effort from executing a task for the robot to teaching it what to notice. The useful system would not need a single universal quality score. It might maintain several judgments, together with the conditions under which each one is relevant.

There is an important difference between evaluating existing behavior and guiding the creation of new behavior. In Scaling Laws for Reward Model Overoptimization, stronger optimization of a proxy reward could worsen performance according to a separate reference reward. The study used a synthetic setup in which a fixed reward model stood in for human judgment, so it was not a direct measurement of every way human preference can fail. It nevertheless isolates a useful concern: a predictor can look adequate on familiar examples and become unreliable when an optimizer deliberately searches for outputs that receive high scores.

For a robot, imagine a visual critic that rewards a tidy table. A policy might produce the expected arrangement while placing fragile objects in unstable positions that the camera does not reveal. This is a hypothetical failure, not a claim about an existing product. It illustrates why agreement on ordinary examples is not enough. We would also need to test the evaluator on behaviors selected by the system itself, on unfamiliar objects, and on outcomes that differ in properties outside the image. A learned evaluator should reduce unnecessary human checking, not silently redefine success as whatever it can see.

This is where qualitative judgment becomes a source of both automation and research. A reliable learned evaluator could make previously expensive adaptation routine. Studying when its judgments travel to new settings, when they fail under optimization, and which additional observations repair them could produce new knowledge. I would not expect human evaluation to disappear, but I would expect its placement to change. Instead of assigning a score to every attempt, people may help identify missing criteria, resolve important disagreements, or check new classes of behavior. Whether that actually reduces their workload must be measured rather than assumed.

The person is also part of the task

There is another limit to treating humans as a source of labels: the person who evaluates a result may not be the only person affected by it. Someone purchases the robot, someone works beside it, and someone else may receive its service. Their priorities need not be identical. Predicting an average rating does not decide whose preference should control a shared task. Nor does making one operator comfortable establish that the robot is physically safe. A system needs to distinguish the authority to specify an outcome from the ability to predict how somebody will rate it.

Even with one user, the human is not always a passive observer. Imagine a robot moving a box across a shared table. If its direction is clear, a person may make room before it arrives. If its intention is unclear, the person may stop working or move an object into its path. The resulting completion time depends on both participants. The robot has partly created the behavior it later measures. This is why I would be cautious about interpreting every improvement during repeated use as the robot becoming more capable. The person may have learned how to accommodate its limitations.

I expect learning systems to become better at choosing when to ask, observe, or proceed with a small reversible action. An uncertain object identity might be resolved by another camera view. An uncertain destination might require a question. Uncertainty about the acceptable appearance of a fold might be easier to resolve through two examples than a long conversation. Asking more questions is not automatically more respectful, and asking fewer is not automatically more intelligent. The useful choice depends on what information is missing, how costly an incorrect action would be, and how much effort the interruption creates.

Correction will also need a scope. If I say, “Please move this cup more slowly,” I may mean this fragile cup, this handover, or all transfers while I am carrying something. A model that remembers the sentence but applies it everywhere has not necessarily learned the intended lesson. A model that discards it after one attempt may be equally unhelpful. I expect systems that retain conditional preferences and make their application inspectable to become more useful than systems that merely collect more examples. Users should be able to correct the interpretation without having to reconstruct the entire policy.

Graphics suggests a richer interface than a single text box. An animator can identify a frame, select a region, constrain a joint, or preserve one part of a motion while changing another. In robotics, I would expect analogous ways to point to a placement, show an acceptable alternative, revise one stage, or mark something that must remain unchanged. This is a design forecast, not a claim that physical action can inherit every affordance of animation. A virtual edit can be undone in ways a physical spill cannot. The interface must account for that difference rather than merely provide more controls.

The generalization of such corrections is an especially important research question. Can a robot transfer a placement preference to another object while preserving contact constraints? Can it change timing without disturbing a successful grasp? Can a user understand what the system changed well enough to stop an inappropriate update? A demonstration that one correction worked once would not answer these questions. I would want to know which corrections remain local, which transfer, and which damage previously reliable behavior. Making robots easier to change will require representations and evaluations that expose these relationships.

When quality becomes ordinary engineering

If these capabilities develop, many forms of customization may become standard product work. Storing a preferred handover location, connecting a validated evaluator, or applying an established adaptation procedure could be ordinary engineering. This would not mean that the work is unskilled or unimportant. It would mean that the main challenge has shifted toward delivering and maintaining a known capability. A reliable tool that lets people obtain the result they want can be the successful outcome of earlier research. It does not need to remain a novel contribution every time someone uses it.

Skinning makes this distinction easier to see. Adjusting one character’s settings is different from discovering a representation that makes whole classes of deformations easier to produce. The latter can be research even when the improvement is expressed through artistic control and authoring effort rather than a new physical capability. Direct Delta Mush is useful here because efficiency, deformation quality, and control were addressed together. Robotics could follow a similar path: setting one user’s preference may become routine, while understanding how preferences can be changed without breaking other behavior remains a substantive method problem.

There is also research that does not produce a universally better algorithm. A careful deployment study might reveal that a supposedly helpful correction mechanism transfers work from the operator to a colleague. An interface might make one kind of uncertainty visible while hiding another. A study could show that an existing quality measure fails for a particular activity or group of users. These are possible research contributions when the observations, interpretation, and scope are established carefully. The value lies in what others can learn from the finding, not in whether the paper can claim an improvement for every possible user.

This means that evaluation has to follow the claim. A system described as easier to correct should help users reach their intended results, not merely require fewer clicks. A system described as more autonomous should not depend on someone quietly preparing every object or watching every attempt. A personalization method should be compared with a strong, simpler alternative, not only an unadapted robot. Time, attention, successful outcomes, unwanted changes, and the distribution of effort may all matter. Some can be counted directly; others require observing and interpreting what people actually do. Neither form of evidence replaces the other automatically.

I do not think this creates a permanent safe territory for researchers in graphics or HCI. Models may improve at interpreting corrections, designing interfaces, and analyzing behavior too. The durable question is not whether a human had to make the artifact. It is whether the work establishes something we did not already understand or makes a previously inaccessible form of investigation possible. That may be an algorithm, a representation, a situated finding, or an evaluation that reveals a consequential failure. The method used to obtain it does not decide its significance in advance.

The next robot may need to be easier to change

Within task families where reliable execution becomes common, I expect the differences between robots to become more visible in use. A user may prefer a system that performs fewer tasks but accepts a correction precisely, explains an interruption through its behavior, and resumes without requiring a complete restart. This does not make broad capabilities irrelevant. It means breadth, reliability, and the effort of obtaining a particular result will form different dimensions of usefulness. A benchmark that records only whether a task eventually finished may increasingly fail to predict which system people can comfortably depend on.

I also expect more of the teaching process to happen through examples, comparisons, and bounded edits rather than complete demonstrations. Shared policies and learned evaluators could supply much of the background knowledge. Local interaction would specify what differs here. The difficult part would be keeping the change appropriately local while allowing useful transfer. A robot that learns every preference from scratch is burdensome. A robot that applies every preference everywhere is intrusive and unreliable. The opportunity is not unlimited personalization but a better relationship between reusable competence and context-specific control.

A related change may occur in what organizations purchase and measure. Instead of valuing only a robot’s catalog of demonstrations, they may care about the work required to introduce another variant, recover from an exception, or hand control to a new operator. I would expect tools for previewing behavior, checking a proposed update, and recovering known settings to matter more as adaptation becomes frequent. Their importance would come from the changing use of robots, not from making every user into a robotics engineer. The best interface may conceal most internal choices while exposing the few that the person needs to understand and influence. More controls are not automatically more agency. Delegating a routine decision can be useful control too, provided the user can recognize and revise an unwanted result.

These forecasts could be wrong in useful ways. Stronger models might infer local expectations so well that elaborate correction interfaces become unnecessary. Good defaults could matter more than extensive control. Conversely, preference learning might work on curated examples yet fail to reduce effort in everyday collaboration. To distinguish these futures, I would study unfamiliar users and tasks over repeated use, including what happens when preferences change. If satisfaction rises only because people learn to avoid the robot’s weak points, that is a different achievement from a system that adapts to their work.

The lesson I take from graphics is not that robotics will eventually become a matter of taste. It is that making a capability available changes the questions we can ask of it. Once a character can move, we can investigate how to make its movement expressive, controllable, and practical to author. As particular robot capabilities become dependable, we may be able to investigate their use with a depth that fragile prototypes have made difficult. Some of yesterday’s research will become ordinary engineering. That can be a sign that it has succeeded. The next research question is what we can now understand and make possible without asking every person to become an expert in the machinery beneath it.