MIT and NYU Unveil FLEX Robot Planning System
The FLEX pipeline lets robots infer higher-level intentions from natural language commands beyond their literal wording.
Updated on Nov. 9, 2026 in Robotics

Researchers from MIT and New York University introduced FLEX, a flexible feature-editing pipeline for robot motion planning. In simulations of a robot arm, FLEX followed literal commands while reasoning at a higher level across three tasks.
Why it matters
Existing methods require users to specify every feature to avoid, which becomes tiresome as the feature list grows. FLEX aims to reduce that burden by letting robots generalize from natural language instructions.
The Generalize step uses GPT-4o to search training data for patterns among specified features, while the Mask step uses computer vision to delete objects no longer present. The full three-step process runs in about five seconds per command.
The players
Massachusetts Institute of Technology (MIT)
A private research university in Cambridge, Massachusetts, known for its engineering and robotics research programs.
New York University (NYU)
A private research university based in New York City with a major presence in computer science and robotics research.
The details
FLEX works in three steps: Add, Generalize, and Mask. The Generalize step uses a large language model to find patterns among features and avoid similar objects, and the Mask step removes features for objects the robot's vision system no longer detects.
Timeline
- November 9-12, 2026: The study is scheduled for presentation at the Conference on Robot Learning (CoRL 2026) in Austin, Texas.
The Tech Race
FLEX applies GPT-4o to robot motion planning, a field that has traditionally relied on mathematical formulations consumers cannot be expected to learn. The approach shows how general-purpose language models are being layered onto specialized robotics stacks.
If the approach matures, users could give robots everyday instructions without listing every object to avoid. That would reduce the setup work needed before a robot can carry out common household or workplace tasks.
The takeaway
Robots that understand intent rather than only literal wording could make everyday commands far less tedious to specify. Watch for physical-hardware results beyond the simulations reported so far.
What happens next
The study will be presented at the Conference on Robot Learning (CoRL 2026) in Austin, Texas, November 9-12, 2026.
Further reading
For more on how language models are reshaping machine control, explore Robotics coverage.






