When we look at the historical development of artificial intelligence, it was revealed how slowly the models that worked under certain conditions developed and how costly they were in the past. Over time, when neural networks developed sufficiently, they therefore replaced old-school developed systems and led to the emergence of large language models. A similar change may occur for robots.
Massachusetts Institute of Technology (MIT) researchers have introduced a method that resembles the working logic of large language models, which can be used instead of robot development methods that can only work under certain conditions. In this method, robots are trained with data from different sources and data they collect themselves. This makes it easier for them to learn a new job or adapt to new conditions.
Inspired by language models
The researchers stated that they were inspired by large language models such as ChatGPT when developing this method. Considering that large language models can be trained with large-scale language data and then used with very little training, the researchers developed a model called “Heterogeneous Pre-Trained Transformers” (HPT). Thanks to this model, data from different cameras and sensors can be processed together, allowing robots to learn new tasks more easily.
Thanks to the new method, robots can act more skillfully and perform different tasks. Researchers will now examine the development potential of this method. They explain their goal as robot brains that can be downloaded and installed like downloading an application.
62847
