MIT Unveils Novel Method of Training General-Purpose Robots Using Generative AI Techniques

Massachusetts Institute of Technology (MIT) unveiled a brand new methodology to coach robots final week that makes use of generative synthetic intelligence (AI) fashions. The new method depends on combining information throughout completely different domains and modalities and unifying them right into a shared language which may then be processed by massive language fashions (LLMs). MIT researchers declare that this methodology can provide rise to general-purpose robots that may deal with a variety of duties without having to individually practice every talent from scratch.

MIT Researchers Develop AI-Inspired Technique to Train Robots

In a newsroom post, MIT detailed the novel methodology to coach robots. Currently, instructing a sure activity to a robotic is a troublesome proposition as a considerable amount of simulation and real-world information is required. This is critical as a result of if the robotic doesn’t perceive the way to carry out the duty in a given surroundings, it’ll battle to adapt to it.

This means for each new activity, new units of information comprising each simulation and real-world state of affairs are wanted. The robotic then undergoes a coaching interval the place the actions are optimised and errors and glitches are eliminated. As a outcome, robots are typically skilled on a selected activity, and people multi-purpose robots seen in science fiction films, haven’t been seen in actuality.

However, a brand new method developed by researchers at MIT claims to bypass this problem. In a paper printed within the pre-print on-line journal arXIv (notice: it isn’t peer-reviewed), the scientists highlighted that generative AI can help with this downside.

For this, information throughout completely different domains, corresponding to simulations and actual robots, and completely different modalities corresponding to imaginative and prescient sensors and robotic arm place encoders, have been unified right into a shared language that may be processed by an AI mannequin. A brand new structure dubbed Heterogeneous Pretrained Transformers (HPT) was additionally developed to unify the info.

Interestingly, the lead creator of the examine, Lirui Wang, {an electrical} engineering and laptop science (EECS) graduate scholar, mentioned that the inspiration for this method was drawn from AI fashions corresponding to OpenAI’s GPT-4.

The researchers added an LLM mannequin known as a transformer (much like the GPT structure) in the midst of their system and it processes each imaginative and prescient and proprioception (sense of self-movement, pressure, and place) inputs.

The MIT researchers state that this new methodology may very well be sooner and cheaper to coach robots in comparison with the normal strategies. This is essentially as a result of lesser quantity of task-specific information required to coach the robotic in numerous duties. Further, the examine discovered that this methodology outperformed coaching from scratch by greater than 20 p.c in each simulation and real-world experiments.