Boston Dynamics, the American company known for its cutting-edge work in robotics, has just formalized a collaboration with Toyota – or more precisely with its subsidiary Toyota Research Institute. Their objective: to accelerate the development of general humanoid robots, capable of carrying out a large number of tasks to serve humans in the flesh.
To achieve this, the two partners will start by working on a first-class guinea pig: Atlas, Boston Dynamics’ most advanced android which has progressed at a breathtaking pace in the space of a few years. We remember in particular his incredible sequences of acrobatics which have no equivalent in this industry.
Twitter-tweet” data-width=”500″ data-dnt=”true”>
It’s time for Atlas to pick up a new set of skills and get hands on. pic.twitter.com/osOWiiBlSh
— Boston Dynamics (@BostonDynamics) January 18, 2023
Recently, the company put aside the old model built around a hydraulic system to focus on a fully electric versioneven lighter and more mobile. Even if Boston Dynamics has remained relatively discreet about its progress since this major shift, this evolution marks the beginning of a new phase of progress, both on the mechanical side and in terms of the software.
AI applied to robots is gaining momentum
This is where Toyota Research Institute comes in, a subsidiary of the Japanese manufacturer which specializes in research applied to AI, robotics, autonomous vehicles, and even materials science. In particular, it claims the status of world leader in the development of what it calls Great Role Models, or LBM (Large Behaviour Model). This term created by analogy with LLMs (the large language models that underpin AI chatbots like ChatGPT) refers to generative AI models specially designed to allow robots to move up a gear in terms of autonomy.
This is an area of research that has remained rather discreet until now – a surprising finding in the current context, where many types of AI models are progressing very quickly. According to TRI director Gill Pratt, this is largely because it has traditionally been difficult to reconcile the two disciplines, due to the inherent limitations of training these systems. “ In machine learning (applied to robotics), until very recently, we had to make compromises “, he explained in September in an interview with TechCrunch. “ The approach works, but it takes millions of scenarios to train the model. Now, when you’re working with physical material, you don’t have time to get that much; the machine breaks down before reaching 10,000. »
But recently, the tide has started to turn. Engineers have gradually moved away from this approach, which is more or less based on brute force, and now favor quality over quantity. By exploiting a small number of examples carefully selected for their diversity, they can now train an LBM with a few dozen reference movements.
The dawn of a new era for Atlas
However, it just so happens that Atlas is already an extremely versatile machine on a purely mechanical level. The two partners will therefore be able to assign lots of new and varied tasks to it in order to collect excellent quality data. These will be used in addition to new simulation systems to train cutting-edge LBMs.
Ultimately, Boston Dynamics and TRI hope to demonstrate that these pre-trained models can pave the way for “ rapid acquisition of new skills “. In practice, this will allow the robot to become increasingly robust and dexterous over time without human intervention. The team will also carry out joint research work to answer “ fundamental questions about training humanoid robots and their interactions with humans ».
It will therefore be appropriate to follow this collaboration closely, because knowing the pedigree of the two partners, we can expect quite spectacular results in the relatively near future. Other general androids with a commercial vocation, like Tesla’s Optimus, better watch out!
🟣 To not miss any news on the WorldOfSoftware, , .
