FLEX Teaches Robots to Generalize From Human Corrections

MIT and New York University researchers have developed FLEX, a system that uses conversational corrections to revise the environmental features guiding a robot’s movements. MIT announced the work on November 9, 2026, ahead of its presentation at the Conference on Robot Learning in November 2026.

FLEX addresses a failure mode in language-guided robotics: telling a coffee-carrying robot not to spill on a laptop can add that object to its obstacle list without capturing why it matters. The system infers a broader preference for avoiding electronics, so the simulated arm adjusts its path around both the laptop and a nearby phone without a separate phone-specific instruction.

Editing What the Robot Tracks

The paper calls FLEX the Flexible Feature Editing Pipeline. It separates feature specification from reward optimization: language corrections identify which measurable properties of a scene matter, while demonstrations determine how strongly those properties should influence the trajectory. That division targets representations that are incomplete, over-specified with irrelevant features, or abstracted at the wrong level.

Each correction passes through ADD, GENERALIZE and MASK. ADD checks whether existing features explain the correction; if they do not, a language model proposes a missing feature and generates code to compute it. GENERALIZE replaces an overly specific feature with one tied to the inferred intent. In MIT’s bowling-ball example, a mug-specific constraint becomes avoidance of fragile objects. MASK removes features that are no longer relevant.

Running MASK last means relevance decisions apply to features after any generalization. The experiments used GPT-4o, and MIT reports that the full three-step interaction takes about five seconds.

Andreea Bobu, who led the study and heads MIT’s Collaborative Learning and Autonomy Research Lab, connected that design to household use: “Everyone’s homes are different, and everyone cares about different things, so no robot comes out of the box knowing what matters to you. What excites me about FLEX is that it can pick that up from a sentence or two instead of asking you to spell out every rule — and that’s the difference between a robot people put up with and one they actually want in their home.”

Simulation Results and Human Ratings

The researchers tested FLEX in 9-by-9 navigation grids and simulated seven-degree-of-freedom Franka tabletop tasks involving coffee, a lit candle and a bowling ball. The arm scenarios used five-stage correction sequences, changing object arrangements between stages so that a previously useful feature could become irrelevant or a hidden misalignment could surface.

Comparisons covered add-only ALGAE, mask-only Mask-From-Exhaustive (MFE), and an unscaffolded baseline that asks the language model for a complete updated feature set in one query. Their normalized reward metric assigns 1 to optimal behavior and 0 to the worst-case trajectory.

The paper reports that FLEX achieved the highest average normalized reward in all seven navigation scenarios. In the scenario requiring all three operations, FLEX averaged 0.99, compared with 0.69 for ALGAE, 0.83 for MFE and 0.78 for the unscaffolded baseline.

In the sequential arm simulations, the authors report that FLEX generally learned the target reward after one or two corrections and maintained performance on earlier and later configurations. The comparison methods required more corrections and degraded as irrelevant features accumulated or their representations overfit individual stages.

A separate in-person study involved 12 participants, each providing typed corrections across three five-stage blocks using FLEX, ALGAE and MFE. The researchers report that FLEX received the highest average rating on all five questionnaire dimensions. For intent understanding on a seven-point scale, the averages were 6.3 for FLEX, 5.0 for ALGAE and 4.5 for MFE. FLEX’s differences from both baselines were statistically significant after adjustment for multiple comparisons. Differences in ease of use and collaborativity were not statistically significant.

Evidence Limits and Development Status

The human study did not run closed-loop on a physical robot. Participants watched simulated trajectories because some baseline methods produced erratic arm motions; the team separately combined their corrections with expert demonstrations collected on a physical Franka for offline learning. The authors describe that hardware work as a feasibility check supporting the simulation study.

FLEX is currently limited to quasi-static manipulation of rigid objects whose relevant properties can be expressed in language; dynamic tasks and deformable objects may require additional modalities. The authors say language-model commonsense can accommodate widely shared preferences but fail for personal or counterintuitive ones. A request to stay even farther from a laptop also encodes preference strength, which the current formulation does not capture.

The authors’ project page is still marked as a work in progress and lists code and dataset availability as coming soon.

For household and assistive robots, fewer corrections are useful only when the inferred preference matches what the user actually cares about. The central open question is whether language-model generalizations can accommodate idiosyncratic preferences with the same consistency during closed-loop operation on physical robots.