Researchers from Microsoft recently published a paper that provides a practical example of how ChatGPT may be used in a few-shot scenario to convert natural language orders into a set of actions that a robot can perform. The prompts were designed with the intention of satisfying the requirements characteristic of many real-world applications while also being prepared for simple adaptability.
Researchers created input prompts to urge ChatGPT to: output a sequence of predetermined robot operations with justifications in a readable JSON format in order to satisfy these requirements; formalize the operational environment’s representation; and infer and output the updated operating environment state, which may then be used as the subsequent input.
This will enable ChatGPT to run entirely off of the memory of the most recent operations. Researchers tested the efficiency of their suggested prompts in diverse situations to infer the right behaviors for multi-stage linguistic instructions through studies.
For this study, they listed the following specifications: simple communication with robot execution systems or software for image recognition; applicability in many domestic contexts; and the ability to convey any quantity of instructions in plain English while lessening the impact of ChatGPT’s token restriction.
Researchers found out that ChatGPT’s conversational characteristics let users alter its output using natural language feedback, which is crucial for developing an application that is both secure and resilient while still providing a user-friendly interface.
The purpose of this study is to provide the robotics research community with helpful information by developing and disseminating generic prompts that are easily adaptable to any experimenter’s needs. They are available on GitHub for free, along with usage instructions, and are open-source.