© Alex Knight/Pexels
Robotics, like AI, are two booming fields. On the one hand, we have increasingly capable robots: Figure 01 which uses the LLM from Open AI, Atlas, the acrobat from Boston Dynamics or even robots killers of Onyx Industries. On the other, chatbots powered with ever more advanced language models like Claude 3 or the future ChatGPT-5. The synergy between these two sectors could completely overturn certain areas: industry, personal assistance, logistics or agriculture.
The foundations of AI for robotics
Tech giants like OpenAI and Google DeepMind are now dedicated to incorporate advanced learning algorithms into the very heart of robots. The latter, commonly referred to as foundation models , give robots the ability to acquire general knowledge and, consequently, to develop their skills. easily adapt to a wide range of tasks. For Gerard Andrews, a marketing manager at Nvidia, the winds of change are blowing. “ We have the deep conviction that we are on the threshold of a radical transformation in robotics ”.
These foundation models exploit neural networks mimicking the structure of the human brain. They are endowed with a unique ability: that of learning from vast heterogeneous sets of data, encompassing texts and images . This immense amount of information allows them to gain deep contextual understanding. As an illustration, Gr00t, the general model launched by Nvidia last March, designed specifically for humanoid robots, demonstrates a quite astonishing willingness to learn and adapt.
Google DeepMind, for its part, has developed an advanced model called Robotic Transformer 2 (RT-2). This is capable of maneuvering a mobile robotic arm manufactured by its sister company, Everyday Robots. By drawing on data gleaned from the Internet and from videos of robots in action, RT-2 is able to follow instructions, even for tasks it has never before been confronted.
Enriching AI with physical experience
The tangible experience of robots proves fundamental to enrich the cognitive faculties of artificial intelligence. According to Akshara Rai, AI researcher at Meta, “ The final step towards true intelligence undoubtedly lies in physical intelligence& ;nbsp;”. An opinion close to that of the CEO of Huawei.
In other words, the physical interactions of robots with their immediate environment are necessaryto develop truly general and efficient artificial intelligence. The integration of physical data allows robots to significantly refine their understanding and ability to adapt to varied environments, thus giving AI increased robustness and versatility. A quid pro quo, in short.
Boston Dynamics perfectly illustrates this approach by deploying its robots like Atlas, which, thanks to a detailed mapping of its environment, can perform very complex complex movements: parkour, object transport, dancing, etc.
AI and robotics: a future yet to be built
Despite the immense progress made recently, there is still a long way to go. Harold Soh, a specialist in human-robot interactions, expresses skepticism about the immediate impact of these technologies. According to him, foundation models for robotics must certainly be explored, but he remains cautious about the so-called revolution predicted by some enthusiastic researchers. Obstacles to overcome include the need for vast amounts of specific data for robot learning, as well as legitimate concerns about the safety and reliability of robotic systems.
Furthermore, researchers strongly emphasize that foundation models require massive and diverse data sets to effectively train robots. Exactly, at the moment, there is no sufficient data set as can be seen for chatbots.
Will AI revolutionize robotics? The prospect is attractive, but we will still need a good dose of patience. Certainly, the demonstrations of certain robots are undoubtedly very impressive, but it should be kept in mind that they are ultimately just simple demonstrations. Large-scale operational deployment of these technologies still remains a fairly distant reality. A reality to which we are nevertheless getting closer and closer every day thanks to the efforts of researchers and other professionals in these sectors.
- < li>Currently, some tech giants are developing AI models specifically designed to integrate into robots: foundation models.
- The physical link between the robot and its environment is what will later give AI a completely different level of efficiency.
- However, the scope for progress is still quite broad before we can imagine and see the large-scale development of robots with true intelligence.
📍 To not miss any Presse-citron news, follow us on Google News and WhatsApp.
[ ]