A crowdsourced phrasebook for robots that read ChatGPT

What it does

PromptCraft-Robotics is a community archive and a testbed for using large language models to direct robots. It hosts curated prompt examples—submitted via GitHub Discussions and voted on by the community—for tasks like aerial inspection, object manipulation, and visual navigation. The repository also bundles a ready-made AirSim simulator environment where a drone takes natural-language commands relayed through ChatGPT.

The interesting bit

The project treats prompt engineering as a communal, empirical science for physical systems rather than a solo hack. Instead of hiding failed attempts, it explicitly welcomes “incorrect” or suboptimal prompts if they reveal something interesting about how LLMs reason—or fail to reason—about physics and spatial relationships. It is essentially a shared lab notebook for robot hallucinations.

Key highlights

• Community-driven prompt curation via GitHub Discussions with upvoting and admin review.
• Covers both simulated (AirSim, Habitat) and real-world hardware (Tello drone, physical manipulation arms).
• Includes a downloadable AirSim binary with pre-built inspection scenes (turbines, solar panels, towers) and a ChatGPT-linked drone interface.
• Backed by a Microsoft Research technical report on design principles and model abilities.
• Prompts span categories from embodied navigation to spatial-temporal reasoning (e.g., visual servoing with a basketball).

Caveats

• Heavily tilted toward OpenAI’s ChatGPT; support for other LLMs is explicitly welcome but still nascent.
• The bundled AirSim environment is distributed as a binary, which suggests limited ability to modify underlying simulation assets.

Verdict

Robotics researchers and prompt engineers who want to see how LLMs translate prose into trajectories will find the curated examples and simulator useful. If you are looking for a drop-in, model-agnostic robot control framework with no API dependencies, this is not it.