Microsoft tests ChatGPT in robots so they can interact with humans

24 Feb. (Portaltic/EP) –

Microsoft has announced that it is investigating how human-robot interactions occur, using ChatGPT artificial intelligence (AI) technology on these machines.

Developed by OpenAI, ChatGPT is a ‘chatbot’ trained to have a text conversation that is based on the GPT 3.5 language model and has the ability to generate and link ideas, as well as remember conversations.

In the last month, the company has been integrating this technology into its different services, starting with the Azure OpenAI Service in mid-January, to later bring it to its Bing search and your Edge browser.

Its latest iteration has materialized in the preview of the new AI-powered Bing on mobile apps and its Edge browser, as well as the search engine version for Skype..

Now, the technological manufacturer has advanced who is testing this technology on robots, to “make natural interactions possible” with humans and has explained what he expects from this integration.

“Our goal with this research is to see if ChatGPT can think beyond text and reason about the physical world to help with robotics tasks,” Microsoft said in a statement, adding that it seeks “help people to interact with robots more easily, without the need to learn complex programming languages ​​or details about robotic systems.”

In this sense, he has established as a challenge to teach this ‘chatbot’ to solve problems taking into account factors such as the context of the operating environment and the way in which “physical actions can change the state of the world”.

The company has recognized that this OpenAI product has capabilities to perform tasks on its own, but that has developed a number of design principles that can be used to guide language models towards solving robotics tasks.

On the other hand, it has indicated that ChatGPT can “unlock a new paradigm of robotics”, in which slow, expensive and inefficient processes of rewriting the code of the robot’s behavior usually follow one another.

In this way, this ‘chatbot’ has the ability to provide high-level feedback to the so-called large language model (LLM), while monitoring the performance of the robot and generating code to different robotics scenarios.

To drive the LLM, Microsoft has built a series of design principles intended to write directions for robotics tasks. To do this, he has defined a set of robotic application programming interfaces (APIs).

Once the text message has been written for ChatGPT to interpret, in which the objective of the task that the machine must perform is described, it is explicitly indicated what it must do, what it must not do or how the ‘chatbot’ You must formulate your answers.

The developer or engineer awaits the response or output of the ChatGPT code using a simulator and, if necessary, uses natural language to assess the AI ​​on the quality and security of its responses.

Once the model has responded in the desired way and the solution is viable, the final code of the artificial intelligence chat is implemented in the robot.

CASES OF CHATGPT IN ROBOTS

The technology company has shared a series of experiments to show what actions robots can respond to using the model developed by OpenAI.

First, he learned the steps to follow to control a real drone and he was able to write complex code structures, such as a zigzag pattern to visualize the shelves of a bookcase.

ChatGPT was also used with the Microsoft AirSim simulator, a scenario in which the model “was able to analyze user intent” and the detailed signals to control a drone with precision.

This language was even tested in a manipulation scenario with a robotic arm. After analyzing the guidelines, ChatGPT was able to link a series of skills learned to, among other activities, stack wooden blocks.

The developers also chose to give these robots objectives such as find Objects, through detection functions. Likewise, he carried out additional experiments to evaluate if the ‘chatbot’ can decide where to direct the machine based on what it should look for.