AI Platform Helps Robots Reason Beyond Literal Commands...

Tell a robot "don't spill on the laptop" and most systems will do exactly that, once, for that one laptop, on that one table. Ask it to do the same task somewhere else and it has no idea the warning still applies. MIT and New York University researchers want to close that gap. At this month's Conference on Robot Learning, a team led by MIT postdoc Nathan Dennler will present FLEX, a feature-editing pipeline built for robot motion-planning that lets machines take a plain-language instruction and extract the reasoning behind it, not just the command itself.

The setup works through text. A person types instructions to a robotic arm in a simulated environment, and FLEX uses a large language model to figure out what general rule a specific command implies. In one test scenario, a robot carries a cup of coffee across a cluttered table while a human operator can type warnings on the fly.

The system is built to take those warnings and generalize them across the task, rather than treating each one as a one-off fix. The researchers see this as a step toward robots that can work alongside people using the same loose, conversational language humans already use with each other.