Why Your Robot Keeps Misunderstanding You

Why Your Robot Keeps Misunderstanding You
https://www.youtube.com/watch?v=Cf4tpWQogzE
The Robot Learns What You Mean

The Robot Learns What You Mean

MIT and NYU’s FLEX system uses plain-language corrections to edit the task features a robot tracks, while exposing the limits of language-model common sense.

Why a Demonstration Can Hide the Real Goal

Picture a robot carrying coffee across a table. You tell it, “Don’t spill on my laptop.” A person can infer a broader concern: liquid can damage electronics, so the nearby phone may matter too. A robot may instead treat the laptop as one isolated obstacle.

That gap begins with the robot’s task representation—the measurable features it uses to score possible motions. A demonstration shows where the arm moved, but it may not reveal why the person preferred that path. The same trajectory could reflect concern about electronics, the table edge, or another nearby object. If the representation omits the feature the person cares about, changing weights on the existing features cannot fully repair the mistake.

The FLEX project from MIT and New York University separates two parts of this problem. Language revises which features belong in the task representation. Demonstrations then help estimate how much those features should influence the chosen trajectory. The aim is not to make a robot generally understand a household. It is to make a specific kind of representation error easier to identify and edit.

Why One-Off Features Become Brittle

Earlier feature-editing methods can add a new item after a correction. That can solve an immediate omission, but repeatedly adding narrow features can leave a redundant representation that does not transfer well when the scene changes.

The FLEX paper studies three ways a representation can be wrong. It can be incomplete because a relevant feature is missing. It can use the wrong level of abstraction, such as tracking one laptop instead of electronics. Or it can be over-specified because an old feature remains active after it stops mattering.

Those distinctions matter in changing environments. A rule that works in one tabletop arrangement may conflict with another object after the scene is rearranged. The researchers designed FLEX to edit the feature set itself instead of treating every correction as another permanent exception.

How FLEX Edits the Checklist

FLEX applies three operations in order: ADD, GENERALIZE, and MASK. ADD asks whether the current features can explain the user’s correction. If they cannot, a language model proposes a missing feature and generates code to calculate it. In the coffee example, distance from the laptop can become part of the task representation.

GENERALIZE looks for the broader category behind that new feature. A laptop can become electronic devices, allowing the revised rule to apply to a phone or calculator that the user did not name. This is where the method uses relationships learned by GPT-4o from text.

MASK removes a feature that has become irrelevant. If an earlier preference pulls the path toward a bowl that now conflicts with the electronics constraint, the user can tell the system to ignore the bowl. The pipeline adds first, generalizes second, and masks last. A secondary report says the full edit cycle takes about five seconds, but the central result is the structure of the edit, not speed alone.

Language supplies semantic information that a trajectory does not contain. It can connect coffee, a laptop, and the broader risk to electronics. That does not guarantee the inferred category is correct. It gives the planner a candidate representation that can be evaluated and corrected.

What the Controlled Tests Found

The researchers first tested FLEX in seven controlled two-dimensional scenarios covering combinations of missing, overly specific, and irrelevant features. In the condition requiring all three edits, FLEX reached an average normalized reward of 0.99. The add-only ALGAE baseline reached 0.69, the exhaustive-mask MFE baseline reached 0.83, and an unscaffolded language-model update reached 0.78. Normalized reward compares the learned trajectory with predefined best and worst outcomes; it is not a household success rate.

The team then evaluated five-stage sequences for a simulated seven-degree-of-freedom Franka arm. Tasks involved carrying coffee around electronics, a candle around flammable objects, and a bowling ball around fragile objects. As the objects changed between stages, earlier representation errors became visible. FLEX learned the intended reward after one or two corrections and retained stronger performance across configurations than the comparison methods.

A user study involved twelve participants—six women and six men—who completed three blocks of five corrections. Participants watched simulated trajectories and typed natural-language feedback. The closed loop remained in simulation because some baseline methods could produce erratic arm motions. Separately, the researchers used the participants’ corrections with expert demonstrations recorded on a physical Franka robot. That was an offline hardware feasibility check, not a person correcting a moving physical arm in real time.

The participant ratings need precise interpretation. FLEX scored significantly higher than ALGAE for task performance, intent understanding, and robustness to unintended consequences. It scored significantly higher than MFE for intent understanding; its robustness advantage over MFE was marginal rather than significant. FLEX had the highest average ease-of-use and collaborativity ratings, but those differences were not statistically significant.

Where Common Sense Stops Helping

FLEX was tested on quasi-static manipulation of rigid objects. Cloth, deformable food, moving people, and other dynamic settings may require touch, motion, sound, or additional perception. The method also depends on a language model’s common-sense patterns. Those patterns can help with widely shared associations, such as keeping liquid away from electronics, but can fail for personal or counterintuitive preferences.

The paper identifies another boundary: FLEX does not fully handle corrections that mix a feature change with a change in intensity. “Stay even farther away from the laptop” says both what matters and how strongly it should matter, while the method assumes those parts can be treated separately.

These limitations carry physical consequences. A mistaken generalization can redirect an arm. A practical interface would need to reveal the inferred rule, request confirmation when confidence is low, and let the user reverse an edit. Those safeguards are future requirements, not demonstrated features of the current system.

An Editable Misunderstanding

FLEX does not show that robots now know what people mean. It demonstrates a more limited and useful idea: when a robot’s feature representation is incomplete, too narrow, or cluttered with an obsolete feature, ordinary language can help edit that representation.

The architecture is easy to state—add what is missing, generalize what is too specific, and mask what no longer belongs. In controlled tasks, that sequence produced stronger and more durable task representations than simpler updates. The next challenge is to make the inferred rules visible and reversible, then test the full interaction safely on physical hardware. The near-term promise is not a robot with human common sense. It is a robot whose misunderstanding can be inspected and repaired before it becomes a physical mistake.

Stay ahead of the curve! Subscribe for more insights on the latest breakthroughs and innovations.