A new research framework wants robots to admit when they can't actually do what an AI planner just told them to do.
Researchers published iAm.md, a Markdown-based standard and generation framework for robot "skill self-assessment," detailed in an arXiv paper posted October 9, 2026. The project targets what the authors call "grounding failures" - cases where a large language model generates a plausible-sounding robot plan without checking whether the robot and its surroundings can actually support it. iAm.md uses open-vocabulary semantic mapping, combining vision-language object detections and segmentation, to build persistent records of objects in an environment that a robot's planner can reference before acting. The team tested the approach on a simulated TIAGo robot across navigation-and-manipulation tasks.
LLM-driven robot planning has leaned on foundation models to turn plain-language instructions into executable code, but those same models can hallucinate steps that have no basis in what a robot can perceive or physically do. By writing an intermediate, standardized representation that both the planner and the robot can inspect, iAm.md gives the system a structured way to flag "I can't actually do this" before committing to a bad plan, rather than relying on the prompt-engineering patches most robotics labs use today.
It's simulation-only so far, and a markdown file won't stop a robot from misreading a bad segmentation mask - but turning hallucination-checking into a format spec, instead of a vibe, is a sign the field is maturing past "the LLM said so."