MIT and New York University researchers have built FLEX, a system that turns one typed correction, such as a warning about spilling coffee on a laptop, into a broader avoidance rule for a robot, and the accompanying paper shows how often that inference goes wrong: each of its three language-model operations erred in 3.8% to 8.6% of isolated trials. The paper is slated for CoRL 2026, the Conference on Robot Learning, which is scheduled for Nov. 9 to 12 in Austin, Texas.

MIT News describes the work as a way for robots to reason beyond literal commands, but its account reports no reward scores or error rates. The paper does report them, and they point two ways: a clear lead over add-only and mask-only editing in the navigation test that needs all three operations, and a design that makes a language model the arbiter of what a robot protects.

Readers evaluating language-guided robots for households or assistive settings should take away three things: how the pipeline works, what the measured gains are, and the limits the authors themselves list.

Add, Generalize and Mask: GPT-4o Edits Feature List While Demonstrations Set Weights

FLEX treats a robot’s reward as a weighted sum of features, written in the paper as a linear function of trajectory features. Language decides which features exist. Human demonstrations decide how much each one matters, with weights learned through maximum-entropy inverse reinforcement learning, a method that infers a reward function from observed behavior.

When a person gives a correction, GPT-4o, OpenAI’s model released in May 2024, edits the feature list through three operations, in this order:

• Add proposes a new feature that captures the correction and generates code to compute it. The paper describes two model calls: one to name the feature and judge whether existing features already explain the correction, and one to write the code.
• Generalize extracts the intent behind the correction and broadens it, so a laptop becomes phones, calculators and other electronics. The paper describes one call to extract intent, then per-feature calls and generated code for the generalized features.
• Mask asks, once per existing feature, whether the feature still matters given the correction, returning a binary keep-or-drop decision.

Mask’s call count grows with the size of the feature set, and the paper reports no runtime or total call counts. MIT News says the process takes about five seconds, a figure the paper does not confirm.

The two accounts also differ on Mask. MIT News attributes it to a computer-vision step that removes features for objects no longer present. The paper describes GPT-4o scoring each feature’s relevance from the correction and the current feature set, with no visual step described.

Normalized Reward Reaches 0.99 for FLEX Against 0.69 for Add-Only Editing

In a navigation scenario that needs all three operations, FLEX scores a normalized reward of 0.99, plus or minus 0.01 standard error, across five random seeds. Its comparison methods score lower: 0.83 for MFE (mask-only editing), 0.78 for USF (a single unscaffolded language-model query) and 0.69 for ALGAE (add-only editing).

The breakdown by operation explains the gap. ALGAE scores 1.00 on add-only tasks but 0.59 when only generalizing is required. MFE scores 1.00 when only masking is required and 0.95 when only adding is required. USF scores 0.64 on add-only tasks and 0.99 on tasks that need add and mask. The tests used seven 9-by-9 grid navigation scenarios, plus simulated five-stage tasks on a seven-degree-of-freedom Franka arm.

Isolated Operations Err in 3.8% to 8.6% of 105 Trials Each

The paper’s appendix reports a reliability test of each operation alone over 105 trials. Add errs in 8.6% of trials (9 of 105), Generalize in 5.7% (6 of 105) and Mask in 3.8% (4 of 105). The errors count both edits applied wrongly and edits wrongly omitted.

Two failure cases from the five-stage tasks show what those errors look like. In stage three, an earlier bowl-proximity feature became a spurious correlation that pulled the trajectory toward the bowl until the user typed a command to ignore it. In stage four, the robot behaved correctly with only a laptop feature because a phone sat next to the laptop, so a hidden misalignment, a robot that never represented the phone at all, surfaced only in stage five when the laptop was removed.

The authors list further limits: the method assumes quasi-static rigid objects, a language model’s commonsense can fail on personal or counterintuitive preferences, and corrections that change how much a person values a preference are ignored. Over-generalization is handled through prompt wording that asks the model to generalize without creating spurious correlations, rather than through a dedicated error analysis.

Hard Limit or Learned Preference: What Does FLEX Avoidance Guarantee?

In FLEX, avoiding an object is a weighted reward feature, so it trades off against the robot’s other reward features. The paper encodes avoidance as a soft term, not a guaranteed constraint, describes no separate safety layer, speed or force limits, or collision constraints, and makes no safety guarantees. Motion-planning research separates the two: planners can optimize multiple objectives such as obstacle clearance, while collision detection tests configuration validity.

The distinction matters because a language model now sits upstream of that cost. When GPT-4o generalizes a laptop to electronics, or when Mask drops a feature that still matters, the change alters the robot’s planned motion. Language-model outputs that state false information as fact are a failure class called hallucination, which ongoing research tries to mitigate and whose proposed fixes can raise computational cost. The FLEX paper does not discuss adversarial prompts or hallucination.

Adjacent research shows why safety teams scrutinize this layer. University of Pennsylvania researchers reported in 2024 that their RoboPAIR method jailbroke three language-model-controlled systems, with attack success rates that often reached 100%, including a commercial Unitree Go2 robot dog. A King’s College London and Carnegie Mellon study published in November 2025 found every tested model was prone to discrimination, failed critical safety checks and approved at least one command that could cause serious harm. Both studies predate FLEX and tested other systems.

12-Person User Study Favors FLEX on Intent Understanding but Not on Ease of Use

In a within-subjects study of 12 participants, 5 of them novice robot users, people typed corrections and rated FLEX above ALGAE on task performance (p = .035), intent understanding (p = .006) and robustness to unintended consequences (p = .047), using Holm-adjusted Wilcoxon tests. Against MFE, FLEX rated higher on intent understanding (p = .031), while robustness fell short of significance (p = .074). Ease of use and collaborativity showed no significant differences.

The authors state the closed-loop study ran in simulation, supported by hardware feasibility checks, and list a closed-loop user study on physical hardware as future work. They say participants watched behavior in simulation for safety reasons, because some baselines produced erratic arm motions. Offline learning from participants’ real corrections was run separately on a physical Franka arm.

Funding Consortium Counts OpenAI Among Founding Members While FLEX Runs on GPT-4o

MIT News says the MIT Generative AI Impact Consortium partly funded the work, along with a Schwarzman College of Computing program on social and ethical responsibilities of computing. The paper’s acknowledgments credit the consortium for partial support and do not mention OpenAI. The consortium introduced on Feb. 3, 2025, with six founding members: Analog Devices, The Coca-Cola Company, OpenAI, Tata Group, SK Telecom and TWG Global. FLEX’s language model is OpenAI’s GPT-4o. MIT News does not say the funding shaped that choice, and the paper reports no comparison with other models.

Model lifecycle is a practical concern for anyone building on the pipeline. OpenAI retired GPT-4o from ChatGPT on Feb. 13, 2026, though the model stayed available through its API.

MIT News names Andreea Bobu as study lead, with co-authors Audrey Lee (SM ’26), Nathan Dennler and Ilia Sucholutsky of New York University. Bobu is an assistant professor in MIT’s Department of Aeronautics and Astronautics who leads the CLEAR Lab and was appointed with a July 2024 start. Her earlier work, Aligning Robot and Human Representations, argues that a robot’s internal task features should match what people care about. Before MIT, she was a research scientist at The AI Institute and an intern at NVIDIA’s Robotics Lab.

Nathan Dennler told MIT News that FLEX gives people a way to tell a robot, “When people say this, they mean that.” Bobu told MIT News that homes differ and people care about different things, and that she is excited FLEX can pick up a preference from a sentence or two.

Language Corrections Outside Claude’s Policy Rules Still Face Hardware Limits

FLEX uses GPT-4o, so Anthropic’s usage policy does not govern it. That policy takes effect Nov. 12, 2026. Its section on high-risk physical actions covers Claude-driven hardware that could injure people without human approval and requires a qualified person able to stop it, a safe state on intervention or lost connection, and limits enforced by equipment or a controller independent of model output. Consumer and household devices within built-in limits are excluded when the worst outcome is discomfort or inconvenience. The same design logic applies to language-guided robots in general: speed and force limits belong outside the language model.

Other groups route language feedback into safety controllers rather than reward weights alone. Researchers including Santos, Li, Peters, Bansal and Bajcsy updated robot safety representations online from natural-language feedback using a reachability-based controller, with hardware experiments. A separate Text2Interaction project from Thumm, Agia, Pavone and Althoff uses language models to generate task plans, motion preferences and safety controller parameters.

Questions to Ask Vendors of Language-Guided Robots

Three questions follow from the paper. Is avoidance a learned preference or a hard limit? Are speed and force limits enforced outside the language model? Do closed-loop results exist on hardware, and for which model version? The paper answers the first two for FLEX: avoidance is a learned feature, and it describes no separate limit layer. It lists a closed-loop hardware study as future work, which answers the third.

Frequently Asked Questions

Can robots understand vague commands like “don’t touch my laptop”?

FLEX infers a broader category from a single typed correction, such as electronics from a laptop warning, using GPT-4o. In the paper’s navigation scenario that needs all three operations, it scored 0.99 normalized reward against 0.69 for add-only editing, but each operation errs in 3.8% to 8.6% of isolated trials.

How does FLEX learn what a person cares about?

Language sets which features the robot tracks. Demonstrations set how much each feature matters, using maximum-entropy inverse reinforcement learning.

Can a language model make a robot less safe?

A wrong generalization changes the robot’s avoidance behavior, and the paper documents spurious-feature and hidden-misalignment cases in simulation. FLEX treats avoidance as a weighted reward feature, not a guaranteed limit, so independent speed and force limits matter.

Has FLEX been tested on a real robot?

Only partly. The authors ran offline learning from participants’ real corrections on a physical Franka arm, but the closed-loop user study ran in simulation, and they list a closed-loop hardware study as future work.