Has Google decided to work on self-coding robots for us all? In response to human-written directions, Google claims that a new open-source benchmark will enable robots to carry out jobs by developing their own code.
The business has unveiled “Code as Policies” (CAP), which enables language-model-generated programs (LMPs) written in Python code to be generated from prompts given in plain English.
CAP is the successor to PaLM-SayCan, a project that provided a comparable interface for plain English commands to control a physical assistant robot. By enabling machines to write their own code, CAP promises to make it possible to carry out more complicated jobs with greater accuracy.
Self-coding robots for us all?
Google Research Intern Jacky Liang and Research Scientist Andy Zeng discuss the release of CAP in a blog post where they detail the technology’s inspiration and potential implications.
“What if robots could autonomously build their own programming to interact with the world when given instructions from people? […] Current language models are very good at writing both generic and, as we’ve found, code that can control robot behaviors when given natural language commands.
However, it might not be quite a time to abandon your programming laptop just yet. Google’s researchers presented straightforward commands with a similar structure during testing. Test robots were able to “place the blocks in a horizontal line at the top” (of a square boundary) and “draw a 5 cm hexagon around the middle” (of a whiteboard).
The project team admits in the paper that goes along with it, “Code as Policies: Language Model Programs for Embodied Control,” that CAP is currently unable to handle instructions that are very abstract or sophisticated or to comprehend descriptions of trajectories. Additionally, the team’s strategy failed to take into consideration impossible commands being fed through CAP.
Theoretically, Google’s “robot-centric” Python version of LMPs, which is open source, might lead to considerably quicker construction of solutions to these problems. Releases via Github and an interactive demo via Google Colab are also available on the CAP website to demonstrate how robots “write” code in response to orders.