Google is working on robots that can create their own code

3 Nov 2022

Image: © Gorodenkoff/

Google researchers believe natural language processing and AI will allow robots to develop their own code to respond to new instructions.

Google is prototyping a system that lets robots write their own code to respond to instructions and perform tasks.

The tech giant said this is designed to reduce reprogramming policies for each new task, which can be time consuming and requires developers with domain expertise.

It could also make it easier for staff to interact with robots on smart factory floors, being able to interact using simple commands without the need to write complex code.

Google researchers have developed language model programs known as Code as Policies (CaP). This involves a code-writing AI model that can create new code for new instructions.

Large language models are natural language processing (NLP) systems that are trained on a massive volume of data.

“Given natural language instructions, current language models are highly proficient at writing not only generic code but, as we’ve discovered, code that can control robot actions as well,” Google robotics scientists said in a blogpost.

For example, Google researchers have been combining large language models with its Everyday Robots so they can respond better to complex and abstract requests from humans.

The Google team said CaP will allow a single system to perform complex and varied robotic tasks without task-specific training.

Their demonstrations show various examples of robotic arms adapting to new instructions, such as moving blocks around in a square before making the square bigger.

“Our experiments demonstrate that outputting code led to improved generalisation and task performance over directly learning robot tasks and outputting natural language actions,” the researchers explained in the blogpost.

They added that CaP imparts a degree of generalisation thanks to pre-trained language models, without the massive amount of data collection needed for end-to-end robot learning.

The tech giant has released the code on GitHub to let others experiment with this system, while researchers plan to learn more by using CaP.

“By characterising the types of generalisation encountered in code generation problems, we can also study how hierarchical code generation improves generalisation,” Google researchers said.

The amount of code being written by AI is on the rise. GitHub recently made Copilot, its AI-powered coding tech, generally available to developers around the world. Copilot was first revealed in June 2021 as an AI assistant for programmers that essentially does for coding what predictive text does for writing emails.

10 things you need to know direct to your inbox every weekday. Sign up for the Daily Brief, Silicon Republic’s digest of essential sci-tech news.

Leigh Mc Gowran is a journalist with Silicon Republic