OpenAI has formed a team to train artificial intelligence models for robots.

The ChatGPT developer confirmed Forbes’ Thursday report in a statement. The news comes about four years after OpenAI disbanded a team that researched ways of using AI to teach robots new tasks. Forbes reported that several former members of the team still work at the company, which could ease its effort to reenter the market. 

According to a job posting on its website, OpenAI plans to develop a suite of multimodal large language models for robotics use cases. A multimodal model is a neural network that processes not only text but also other types of input. That input can, for example, include data from a robot’s onboard sensors.

A number of other companies are already integrating LLMs into autonomous machines. Covariant, a well-funded startup founded by former OpenAI researchers, recently debuted an LLM called RFM-1 for the robotics market. The model allows robots to ask for pointers from a human when they encounter a challenging task.

RFM-1 also lends itself to other use cases, according to Covariant. Before performing a task such as moving a box, a robot equipped with the LLM can generate a short video that depicts how the action might be performed. The robot can then study this video to find the most efficient way of moving the box.

OpenAI’s job posting didn’t specify how the LLMs that it will build for the robotics market might work. However, the ChatGPT developer detailed that the position’s responsibilities include “exploring new model architectures.” This hints the LLMs the company is developing might be based on a different design than the decoder-only Transformer architecture that underpins most advanced language models.

According to Forbes, OpenAI formed its robotics team about two months ago. The company is expected to make the LLMs that the team will develop available to third-party robot manufacturers. OpenAI has reportedly “gone head to head” with a number of companies in this market as part of its push to recruit engineers who can support the development effort.

OpenAI’s launch of the robotics unit comes four years after it disbanded an internal team that focused on the same market. 

In 2017, the company’s original robotics team detailed an AI system that allowed robots to learn by demonstration. Engineers could teach the system new tasks by demonstrating how those tasks should be performed using a virtual reality headset. The following year, OpenAI released a set of simulation tools for teaching robots’ AI software new skills.

Photo: OpenAI